Robert Vasquez, serving a life sentence for the rape of a child under the age of thirteen, petitioned the district court for a writ of habeas corpus. He argued that he received constitutionally ineffective assistance of counsel because his attorney failed to interview or locate several potential witnesses whose testimony would have cast doubt on the credibility of the victim, the state’s lone witness establishing Vasquez’s guilt. The district court conditionally granted the writ, holding that the Ohio courts unreasonably applied federal law in denying his ineffective assistance of counsel claim. Warden Margaret Bradshaw now appeals from this decision, charging the district court with failing to accord the state courts deference under the Anti-Terrorism and Effective Death Penalty Act and with misapplying Sixth Amendment law to the attorney’s actions. While we agree with the warden that the district court improperly disregarded the factual findings of the Ohio courts, we affirm the district court’s decision. The Ohio courts applied law contrary to clearly established federal law, stating and applying an overly exacting standard for prejudice as a result of deficient counsel performance. On our own review of the record, even in light of the state courts’ factual conclusions, we hold that Vasquez received constitutionally ineffective assistance of counsel and is entitled to habeas relief.
I
A
Vasquez was tried for the rape and kidnapping of nine-year-old A.L. On direct *106appeal, the Ohio courts summarized the facts of the crime, as established by the evidence at trial, as follows:
On July 28, 2000, the nine-year-old victim and her father [S.L.], who is a Cleveland police officer, went to Don Shaffer’s house. Shaffer is the victim’s father’s best friend. The father and the victim met Shaffer and his fiancee Becky Egbertson [now Becky Shaffer] and Becky’s sister, Karra Vasquez, earlier in the evening. They decided that Shaffer would pick up some fast food and they would go to Shaffer’s house to eat. The victim’s father first returned home to drop off his boat.
By the time the victim and her father arrived at Shaffer’s house, it was late and Shaffer and Becky had retired for the night. However, staying with Shaffer was Becky’s sister, Karra Vasquez and her husband, the defendant, Robert Vasquez, and their two small children ....
After they finished [eating], the victim’s father went outside with Kara Vasquez, leaving the victim with Robert Vasquez. Vasquez invited the victim to go downstairs into the basement to watch television and to help him with his two small children. In the basement was a bunk bed. The victim testified that she climbed onto the top bunk and that Vasquez also climbed onto the top bunk after removing the bunk ladder. According to the victim, he then pushed her down with his arm and, while holding his hand over her mouth, he pulled down her pants and underwear and began licking her “private spot.” As he was doing this, his two infant children were crawling around on the bottom bunk. He continued the assault until the victim’s father called to her from the top of the stairs. The victim then put her clothes back on and went home.
The victim did not say anything about the assault to her father at that time. The next day, the father took the victim and her sister to spend a week with their grandparents.... When the girls returned, the father took them to Shaffer’s house to babysit. Sometime later, Robert Vasquez and Don Shaffer returned from work. Robert Vasquez asked the victim several times to go down into the basement with him, but she refused. She then told Don Shaffer that she needed to talk to him. She told Shaffer that Robert Vasquez had sexually assaulted her.
When the victim’s father returned, Shaffer informed the victim’s father what she had told him. The father called his partner at the police department to request that he come over to the house. He and his partner then talked with his daughter to explain the serious nature of the charges and to assure she was telling the truth. He then called 911. During this entire time, Vasquez remained in the basement until he was arrested later that night.
The victim was taken by ambulance to Metro Hospital where she was interviewed by a social worker and a physical exam was performed.
State v. Vasquez, 2001 WL 1352781 at *1-*2 (Ohio Ct.App.2001).
At trial, A.L.’s testimony was the only evidence presented by the state to demonstrate that the crime occurred or that Vasquez committed it. The state also offered testimony from A.L.’s father, Don Shaffer, the investigating police officer, and the social worker who interviewed A.L. Their testimony tended to provide circumstantial evidence of her credibility: that her story remained consistent throughout the investigation and that she acted out of the ordinary, including being fearful for her safety and calling her father daily from *107school. The defense put on no evidence. Vasquez himself did not testify and no other witnesses were subpoenaed or called. Instead, Vasquez’s attorney, Don Butler, pursued a strategy of highlighting the internal inconsistencies in A.L.’s testimony and casting doubt on her truthfulness.
The jury found Vasquez guilty. He was sentenced to life imprisonment for the rape of a minor under the age of thirteen and nine years on the kidnapping charge.
B
In addition to his unsuccessful direct appeal, Vasquez filed a petition for post-conviction relief in the Ohio Court of Common Pleas, alleging he received ineffective assistance of counsel. Because the resulting opinion is not clear as to which of its statements were factual findings and which were descriptions of testimony without credibility determinations, we recount the arguments and evidence presented in more detail than a habeas court ordinarily would.
1
Vasquez’s petition argued, among other things, that his counsel failed to investigate adequately the circumstances surrounding the crime and A.L.’s accusation. This alleged failure resulted in his defense lacking several willing witnesses who could have cast doubt on A.L.’s version of events and on her credibility.
Specifically, Vasquez claimed that Butler, who was assigned to the case by the court after Vasquez’s original court-appointed attorney asked to be relieved, never prepared a defense. The two men met only three times prior to the trial. Vasquez alleges that he indicated to his attorney that his wife, Karra Vasquez, and his mother-in-law, JoAnn Kitchen, could provide helpful information but admitted that he did not know where his wife was staying during the trial. Vasquez also claimed that in addition to the face-to-face discussions, he sent Butler two ten-page letters detailing his version of events, but that Butler never mentioned the letters or followed up on the suggestions in them.
Vasquez also argued that Butler failed to locate available evidence and witnesses who would have challenged AL.’s story and credibility. For instance, the EMS run-sheet produced by the responding EMT on the night that A.L. first accused Vasquez contained statements from A.L. that appear to be at odds with the state’s theory of the crime. The run-sheet indicates that she told the EMT that Vasquez “put it in me” and that she had immediately showered on the night of the assault. At trial, on the other hand, she testified that Vasquez used only his tongue and indicated that she slept through the night on the couch with her father. Vasquez argued that these discrepancies from the initial revelation of the attack would have been a powerful impeachment of her trial story.
Vasquez also identified five witnesses who indicated they would have testified on Vasquez’s behalf had Butler approached them. First, Karra Vasquez said that she called Butler twice but never received a return phone call. Vasquez argued that his wife could have testified to facts that would have made the attack less plausible: her presence in the small house, the basement door being open, the relatively short duration that A.L. was in the basement, and A.L.’s undisturbed clothes and normal demeanor upon returning from the basement. She also would have testified to the circumstances related to the accusation, specifically that she overheard AL.’s father coach A.L. about her story prior to telling the police.
*108Second, Becky Shaffer also indicated that she had not been contacted by Vasquez’s attorney. Vasquez claimed that she had information about the evening of A.L.’s accusation that could cast doubt on its credibility. She rode to the hospital in the ambulance with A.L. that night. She would have testified that A.L. did not appear upset and was excited by the attention, even hoping that her accusation might get her on TV. She also would have testified that A.L.’s father was prompting his daughter’s answers.
Third, Vasquez proffered the testimony of JoAnn Kitchen, Karra and Becky’s mother. Kitchen had no direct knowledge of the crime or of A.L.’s accusation, but said that she was willing to testify as to A.L.’s truthfulness. As a family friend (in addition to the connection between her son-in-law Don Shaffer and A.L.’s father, Kitchen was familiar with A.L. and her family because A.L.’s father had dated a third daughter), Kitchen had frequently babysat for A.L. and interacted with her at family events. In her opinion, A.L. craved attention so much that she could not be trusted to tell the truth. Vasquez also argued that Kitchen was important to his ineffectiveness claim because she was the only family member who actually spoke with Butler. She claimed that in her conversation with Butler, she told him that she could not help, but denies that she refused Butler access to her daughter or indicated that the family was unwilling to help.
Finally, Vasquez offered two additional witnesses: Tammy Salopek and her daughter, Ashley Snyder. Salopek is a family friend to both A.L.’s family and the Shaffers. Vasquez claimed that she was thus easily identified and located by Butler if he had spoken with Becky Shaffer about mutual acquaintances. Both Salopek and her daughter had extensive social contact with A.L. They indicated they would have testified to A.L.’s lack of truthfulness— specifically her willingness to lie to get attention. They also could have testified to A.L.’s demeanor immediately after the accusation, as A.L. slept over at Salopek’s home the following week. Contrary to the evidence at trial, neither mother or daughter observed any strange or out-of-the-ordinary behavior in A.L. Further, Vasquez argued that Salopek and her daughter would have provided more substance to a defense theory that A.L. fabricated the story because Ashley Snyder had, earlier in the summer of 2000, been the victim of sexual molestation and had shared her story with A.L.
Together, Vasquez argued that these witnesses demonstrated that his attorney failed to locate evidence and witnesses that could have undermined the plausibility of the case and A.L.’s credibility. Vasquez pointed out that this testimony tended to show that the layout of the house made the attack unlikely and the fact that she changed her story from the EMS report and received coaching from her father. Moreover, a competent attorney could have put on a defense that explained how a nine-year-old would be familiar enough with sexual assaults to manufacture a story (months earlier she heard details of Ashley Snyder’s own sexual molestation) and why she might do so (she has a history of lying for attention; she even hoped to be on TV after accusing Vasquez). Accordingly, Vasquez asked for a new trial.
2
The Ohio court held an evidentiary hearing on Vasquez’s claims of ineffective representation. The court took testimony from the five witnesses recounted above who supposedly would have participated in Vasquez’s defense but for Butler’s failure *109to locate them during his investigation. Vasquez and Butler also testified.
The five witnesses each testified consistently with their affidavits and Vasquez’s petition described above, outside of a few concessions on cross examination. (For instance, Becky Shaffer admitted that A.L.’s laughing in the ambulance was in response to the EMT’s joking.)
Vasquez testified that prior to trial he had difficulty even learning the identity of his attorney. Once he did discover his attorney’s name and contact information, he met with him only three times, and none of those meetings lasted longer than a few minutes and involved little substantive discussion of his case. He asserted that he identified at least his wife and JoAnn Kitchen as people to contact. He repeated that he sent letters to Butler and that claimed that it appeared Butler ignored them, asking basic questions in their meetings that had been answered in the letters.
Butler denied Vasquez’s characterization of his defense. He admitted that he met with Vasquez only three times, but argued that he discussed the case at length with his client, but that Vasquez provided no helpful information. When Butler did find Karra Vasquez at her mother’s home, he claimed that JoAnn Kitchen refused to speak with him except to indicate that Karra (and the rest of the family) would not help the defense. As to Becky Shaffer, Butler indicated that Don Shaffer’s friendship with A.L.’s father and the fact that he was a state’s witness made him believe it unlikely that she would help the defense. He also indicated that he did research into child sex abuse cases generally and that he concluded that the best strategy was to attack A.L.’s testimony.
After rehearsing this testimony in detail, the court denied Vasquez’s petition. The court explained that “Butler’s testimony not only showed no failure in his obligation to prepare the case for trial in general, but specifically concerning the subpoenaing of witnesses, his testimony and of others showed that the defendant’s family members refused to cooperate with the defense.” Ibid. The court continued: “perhaps most importantly, not one of the witnesses who testified ... offered any testimony that could have changed the outcome of defendant’s trial.”
Reviewing for abuse of discretion, the Ohio Court of Appeals affirmed the denial of Vasquez’s petition. The Supreme Court of Ohio denied discretionary review.
C
Vasquez filed his petition for a writ of habeas corpus with the district court in June 2005. His listed grounds for relief included, inter alia, his assertion that his attorney’s limited investigation was unreasonable and therefore denied him effective assistance of counsel.
The district court first acknowledged that it “must generally defer to the factual finding of the state courts,” but declined to do so here because “[a] comprehensive review of the 800-plus pages of transcript in the record, in addition to the rest of the documents contained in the record, provides clear and convincing evidence that the facts as found in the Direct Appeal Opinion and Post-conviction Hearing Findings ... are inaccurate in numerous places.” The district court, now feeling freed of its usual deference, thereafter undertook a comprehensive discussion of all of the post-conviction testimony, complete with citations to the record.
Turning to the legal question, the district court concluded that “the state court application of Strickland in this case was objectively unreasonable.” Specifically, “Butler’s performance as a whole was defi-*110dent.... Butler simply failed in his duty to conduct an objectively reasonable investigation on which to base his conclusions about trial strategy....” The court highlighted Butler’s failure to identify or interview several potential defense witnesses and the short amount of attention given directly to Vasquez as evidence of the deficiency. And this “deficient performance prejudiced the defense sufficiently to undermine the reliability of the trial, and thus the state appeals court’s findings to the contrary were an objectively unreasonable application of Strickland’s prejudice element.” Accordingly, the district court granted his petition.
II
The Anti-Terrorism and Effective Death Penalty Act (AEDPA), 28 U.S.C. § 2244 et seq., requires us, when adjudicating a petition for habeas corpus challenging the legality of a state court conviction, to defer to a final decision on the merits — of fact or law — of the state court that first decided the claims raised. The same Congressionally-supplied standards governed the district court’s consideration of Vasquez’s petition and, accordingly, our focus is not on the errors that the warden charges to the district court, but on the state court record and whether it warrants the relief Vasquez requests. See Parker v. Renico, 506 F.3d 444, 447 (6th Cir.2007) (“We review de novo the district court’s decision to grant or deny habeas relief.”).
A
AEDPA requires that the writ shall not issue unless we determine that the state court’s adjudication:
(1) resulted in a decision that was contrary to, or involved an unreasonable application of, clearly established Federal law, as determined by the Supreme Court of the United States; or
(2) resulted in a decision that was based on an unreasonable determination of the facts in light of the evidence presented in the State court proceeding.
28 U.S.C. § 2254(d).
For an allegation of constitutionally ineffective assistance of counsel, the relevant “clearly established federal law” is Strickland v. Washington, 466 U.S. 668, 104 S.Ct. 2052, 80 L.Ed.2d 674 (1984). A counsel’s performance is a deprivation of a defendant’s Sixth Amendment right to counsel where a defendant shows his counsel’s assistance was both deficient and that the deficiency prejudiced the defendant. A performance is deficient only if it “fell below an objective standard of reasonableness” in light of the “prevailing professional norms.” Id. at 687-88, 104 S.Ct. 2052. And a deficient performance is prejudicial if “there is a reasonable probability that, but for counsel’s unprofessional errors, the result of the proceeding would have been different. A reasonable probability is a probability sufficient to undermine confidence in the outcome.” Id. at 694, 104 S.Ct. 2052.
The Ohio trial court did not follow this clearly established law. To be sure, the court cited Strickland and identified a two-part test, measuring for deficiency and prejudice. But the court stated that prejudice occurs only when “the result of petitioner’s trial or legal proceeding would have been different had defense counsel provided proper representation.” (emphasis added). After reviewing both Ohio and federal law, the court repeated itself, explaining that “[i]n order to demonstrate a claim of trial counsel’s ineffectiveness according to the United States Supreme Court, a post-conviction petitioner is required to demonstrate that (1) the performance of defense counsel was seriously flawed ... and (2) the result of petitioner’s *111trial or legal proceeding would have been different had defense counsel provided proper representation.” (citing Strickland (without providing a pin cite)). This was not a slip of the pen during a rote repetition of a rule corrected upon application of law to fact. In analyzing Vasquez’s claim as to the failure to interview and call potential defense witnesses, the court relied directly on its version of the ineffectiveness standard as crucial to its decision against relief: “perhaps most importantly, not one of the witnesses who testified [at the post conviction hearing] ... offered any testimony that could have changed the outcome of defendant’s trial.” The court repeated, in the context of a different claim of ineffectiveness, that “even if the Court found that Butler’s question demonstrated a deficiency ... defendant has shown no resulting prejudice. As demonstrated above, neither the testimony of Mrs. Vasquez or Becky Shaffer would not [sic] have changed the outcome of the trial.”
The Ohio Court of Appeals, reviewing for abuse of discretion,1 repeated the error. It framed its rejection of Vasquez’s appeal in terms of a changed-outcome: “defendant presented numerous witnesses who did not testify at his trial. Their absence ... does not demonstrate his counsel was ineffective and that the outcome of his trial would have been different had they been witnesses .... ” (emphasis added). In analyzing the argument, the appeals court expressly adopted the trial court’s reasoning: “[w]e agree with the trial court’s conclusion that ‘not one of the witnesses who testified ... offered any testimony that could have changed the outcome of defendant’s trial.’ ”
This is not a casual error. A “reasonable probability” of difference does not mean “would have been different.” The latter formulation puts a greater burden *112on the petitioner. To prevail on his claim as it was adjudicated, Vasquez was required not only to show that his counsel’s deficiency “undermine[d] confidence in the outcome,” Strickland 466 U.S. at 694, 104 S.Ct. 2052, but to prove that a trial with competent counsel actually would have resulted in his acquittal. In interpreting what Congress meant in § 2254(d), the Supreme Court used a similar mistake as the paradigmatic example of an application of law “contrary to clearly established federal law” that deserves no deference under the statute. The Court explained:
A state court decision will certainly be contrary to our clearly established precedent if the state court applies a rule that contradicts the governing law.... Take, for example, our decision in Strickland .... If a state court were to reject a prisoner’s claim of ineffective assistance of counsel on the grounds that the prisoner had not established by a preponderance of the evidence that the result of his criminal proceeding would have been different, that decision would be ‘diametrically different,’ ‘opposite in character or nature,’ and ‘mutually opposed’ to [the Court’s] clearly established precedent because [it] held in Strickland that the prisoner need only demonstrate a ‘reasonable probability that ... the result of the proceeding would have been different.’
(Terry) Williams v. Taylor, 529 U.S. 362, 405-06, 120 S.Ct. 1495, 146 L.Ed.2d 389 (2000) (internal citations omitted).
Of course, the error there, picked by the Supreme Court for its clarity in illustrating application of contrary law, precisely defined the quantum of proof. The Ohio courts left it at “would have been different,” unadorned by the truly offending words “preponderance of the evidence.” It could be argued that, in the spirit of the deference demanded by AEDPA, we should give the Ohio trial court’s statement the benefit of the doubt and assume that the court implied the words “reasonable probability.” See Holland v. Jackson, 542 U.S. 649, 654-55, 124 S.Ct. 2736, 159 L.Ed.2d 683 (2004) (reversing a decision that the state court decision was “contrary to” Strickland because “use of the unadorned word ‘probably’ is permissible shorthand when the complete Strickland standard is elsewhere recited.”). After all, as the warden argues, the appellate court, consistent with Supreme Court precedent, stated the correct standard once.
We are not persuaded. Different standards make for different outcomes. Even where a state court “explicitly delineated the Strickland test,” its decision is “contrary to” federal law where the court applied an “incorrect burden of proof.” West v. Bell, 550 F.3d 542, 552 (6th Cir.2008). Indeed, our court has already held that a state’s court use of a “would have compelled acquittal” formulation is “contrary to” federal law. Tinsley v. Million, 399 F.3d 796, 807 (6th Cir.2005). While the appellate court did say “reasonable probability” once, the use of the incorrect words cannot be regarded as anodyne “shorthand,” e.g. Woodford v. Visciotti, 537 U.S. 19, 23, 123 S.Ct. 357, 154 L.Ed.2d 279 (2002), because they actually describe and apply a different standard. The court of appeals emphasized the inability to meet the prejudice prong, underscoring whether the trial “would have been different,” and expressly adopted the erroneous legal reasoning of the court below. Accordingly, we hold that the Ohio courts applied law that was contrary to clearly established federal law and we are therefore “unconstrained by § 2254(d)(1) ... and de novo review is appropriate.” Fulcher v. Motley, 444 F.3d 791, 799 (6th Cir.2006) (citing Williams, 529 U.S. at 407, 120 S.Ct. 1495).
*113B
AEDPA also supplies our standard for reviewing a factual record. The statute provides that “a determination of a factual issue made by a State court shall be presumed to be correct.” 28 U.S.C. § 2254(e)(1). Only if Vasquez “rebut[s] the presumption of correctness by clear and convincing evidence” may we disregard those determinations. Ibid. The district court held that this burden was carried. Vasquez agrees; the warden argues strenuously that the district court erred.
Before resolving this dispute, its necessary to clarify what the issue is. The Ohio trial court’s decision was opaque about its findings of fact. Its opinion, to be sure, contained a section labeled “Findings of Fact.” But the section gives no definitive statement of “the facts”. Nor does it resolve any of the testimonial inconsistencies discussed. Instead, it provides the background to its decision: a statement of the applicable legal rules, a summary of the factual testimony given by each witness at the post-conviction hearing, and portions of trial transcript relevant to certain post-conviction issues. None of this is the sort of “determination of a factual issue” that § 2254(e)(1) requires us to accept.
The court’s “Conclusions of Law” section, however, does make a series of statements, which reveal some of the facts that it evidently found essential to its decision. The dispute over whether the district court accorded the proper deference is limited to only these statements, where the district court’s “determination of a factual issue” are actually discernable from its decision. Those statements are:
• “[T]he defendant’s family members refused to cooperate with the defense.” Specifically, “Butler explained, and the witnesses themselves corroborated, that Mrs. Vasquez, Becky Shaffer and Mrs. Kitchen refused to cooperate in his defense”
• “Butler’s explanation concerning his efforts to contact these family members and their uncooperativeness was corroborated by detective Chappelle”
• “Since Tammy Salopek was a friend of Becky’s, had Becky been cooperative with Butler, he could have possible [sic] learned of her, and her daughter’s availability.”
In light of the number of witnesses and the relative complexity of the fact situation, this lack of detail is both surprising and frustrating to a federal court attempting to reconstruct the trial and the post-conviction hearing to decide if habeas relief is warranted. This frustration likely explains the district court’s decision to conduct its own “comprehensive review” of the record and, ultimately, to rely wholesale on that review in place of the Ohio court’s findings, without demonstrating which, if any, findings are specifically rebutted by clear and convincing evidence. But, whether or not understandable, the district court did not do what AEDPA required of it.
Under § 2254(e)(1), a proper comparison of these limited findings of fact to the record asks whether they are contradicted by clear and convincing evidence. The state court’s findings as to Karra Vasquez and Kitchen are not. First, the state court’s conclusion that Kitchen was not cooperative resolves the credibility dispute between Butler (who insisted she refused to help and said no one would) and Kitchen (who denied that she refused to help) in favor of the attorney. A reviewing court, relying on cold transcripts, is not in position to displace such a resolution. See Anderson v. City of Bessemer City, 470 U.S. 564, 574, 105 S.Ct. 1504, 84 L.Ed.2d 518 (1985). Second, the state court’s reference to the corroboration of Butler’s ef*114forts to contact Karra Vasquez means that Butler’s only means of contacting Karra Vasquez was through JoAnn Kitchen. (Detective Chappelle testified to the same difficulty in contacting Karra Vasquez.) If Butler is to be believed over Kitchen, as the state court’s opinion requires, Kitchen’s refusal also blocked his access to Karra Vasquez.
In contrast, the court’s discussions regarding Becky Shaffer is not a determination of fact that forecloses Vasquez’s argument. It is trae that the court grouped her with Karra and Kitchen as someone that “refused to cooperate.” But it is dubious that this statement, contained in a summary statement listing Karra Vasquez and Kitchen, reflects a factual determination by the court. See Wiggins v. Smith, 539 U.S. 510, 530-31, 123 S.Ct. 2527, 156 L.Ed.2d 471 (2003) (evaluating a state court’s conclusory factual finding in light of its explanation and denying it deference under § 2254(e)(1)). Unlike Kitchen (and Karra Vasquez, who was known to be living with Kitchen at the time Butler called her), Butler never contacted or even attempted to contact Becky Shaffer. There was therefore no opportunity for her to “refuse to cooperate” and the court could not have so determined. Instead, its reasoning that “Butler’s professional opinion that Becky Shaffer would not be of assistance since her husband was a State’s witness, and because of Mrs. Kitchen’s representations that no one in the family would help, is a decision that is solely within defense counsel’s discretion” is a better reflection of the court’s holding as to Becky Shaffer. To the extent that the court relied on her “refusal,” it appears from this statement that it was actually deciding that Butler’s conclusion that she would have refused was reasonable. That was a legal conclusion and we review it as such below.
The trial court’s conclusions about the EMS run-sheet also do not merit § 2254(e)(1) deference. The court concluded that the information it contained actually corroborates A.L.’s trial testimony. While credibility determinations as between counsel and potential witnesses regarding what happened pretrial must be deferred to as a factual determination, this conclusion is different. The court decided that one interpretation of the evidence (the EMS run-sheet corroborated A.L.’s testimony) was better than another (that it contradicted both A.L.’s and her father’s testimony). But both would have been permissible interpretations for the jury to draw had the EMS run-sheet been presented to the jury in the first instance, and so this reasoning is actually a holding that Vasquez has not carried his burden to show prejudice resulting from his attorney’s failure to identify the run-sheet as useful impeachment evidence. See Ramonez v. Berghuis, 490 F.3d 482, 490 (6th Cir.2007) (“[W]hat the state court has really done is to state its view that there is not a reasonable probability that the jury would believe the testimony and thus change its verdict.”). Accordingly, we review it de novo below.
To reiterate, on the facts as we must take them, based on the implications of the state court’s decision and the operation of AEDPA deference, Vasquez cannot base his ineffective assistance of counsel argument on the failure to investigate Karra Vasquez or JoAnn Kitchen. They refused to cooperate and any amount of investigation by Butler would not have made a difference. But Vasquez may base his claim on the failure to investigate and put in the record the EMS run sheet and Becky Shaffer’s account of A.L.’s credibility. By extension, Vasquez may also rely on the testimony of Salopek and Snyder, who (by admission of the trial court) likely *115would have been found through an interview with Shaffer, in making his claim.
Ill
The ultimate issue, then, is whether the failure by Vasquez’s counsel to uncover the impeachment evidence from the EMS run-sheet, Becky Shaffer, Tammy Salopek, and Ashley Snyder resulted from deficient performance and, if so, whether it prejudiced the outcome of the trial.
A
Constitutional competence is not a high bar for an attorney to reach. The question is whether the representation “fell below an objective standard of reasonableness.” Strickland, 466 U.S. at 688, 104 S.Ct. 2052. In determinating whether a performance was reasonable, we give a healthy amount of deference to counsel’s tactical and litigation decisions; they are “virtually unchallengeable” in the ordinary case. Id. at 690, 104 S.Ct. 2052. But this deference to strategic choices has always been tempered by a requirement that the choices themselves be informed: “strategic choices made after less than complete investigation are reasonable precisely to the extent that reasonable professional judgments support the limitations on investigation.” Id. at 691, 104 S.Ct. 2052 Accordingly, “a particular decision not to investigate must be directly assessed for reasonableness.” Ibid. Our circuit has expressed this principle with more color: “the investigation leading to the choice of a so-called trial strategy must itself have been reasonably conducted lest the ‘strategic’ choice erected upon it rest on a rotten foundation.” Ramonez, 490 F.3d at 488.
The law as developed so far has not settled on a definitive statement of what a “reasonable investigation” entails. But well established principles suggest that Vasquez’s claim that his attorney failed to identify key evidence and failed to locate and interview critical witnesses is within the known contours of the duty. See, e.g., Towns v. Smith, 395 F.3d 251, 258 (6th Cir.2005) (noting that the duty on counsel to investigate “includes the obligation to investigate all witnesses who may have information concerning his or her client’s guilt or innocence.”); see also Clinkscale v. Carter, 375 F.3d 430, 443 (6th Cir.2004) (collecting cases for the proposition that failing to investigate or call potential defense witnesses constitutes constitutionally deficient representation). That the missing evidence alleged here is impeachment evidence and not, as in some ineffective assistance cases, direct evidence of innocence does not reduce the applicability of this principle. See Tucker v. Ozmint, 350 F.3d 433, 444 (4th Cir.2003) (“Trial counsel have an obligation to investigate possible methods for impeaching a prosecution witness, and failure to do so may constitute ineffective assistance of counsel.”). This is because we are assessing the reasonableness of the representation and the defenses available to the attorney change what is reasonable: where there are no direct witnesses to the alleged crime beyond the perpetrator and the victim, impeachment evidence is at a premium.
The circumstances ordinarily surrounding an accusation of child sexual abuse underscore this concern for developing impeachment evidence. As the Second Circuit has explained, “these cases frequently hinge on judgments about credibility in which jurors must choose between contradictory stories proffered by the defendants and the complainants” because “third-party witnesses [are] often unavailable.... ” Eze v. Senkowski, 321 F.3d 110, 112 (2d Cir.2003). Thus, that court adopted a rule that “defense counsel is obliged, wherever possible, to elucidate any inconsistences in the complainant’s testimony, protect the *116defendant’s credibility, and attack vigorously the reliability of any physical evidence of sexual contact between the defendant and the complainant.” Ibid. While adopting a unique standard of care for sexual abuse cases is unnecessary at this time, the Second Circuit’s admonition is compelling here, in light of what Butler knew as he was undertaking his investigation and defense of Vasquez.
Prior to trial, Butler must have known there was no plausible defense of Vasquez other than attacking A.L.’s credibility. The only witnesses to the crime were A.L. and Vasquez, so producing a witness with a different account was impossible. There was no physical evidence for him to counter, or to demonstrate that a different person was involved. There was no possibility of an alibi; Vasquez was undeniably in the basement with A.L. And in fact, Butler identified no defense witnesses and called none at trial.2
Indeed, viewing the situation (as we must) “from counsel’s perspective at the time,” Strickland, 466 U.S. at 689, 104 S.Ct. 2052, four facts were clear and should have pointed him toward the investigation necessary to put on a stronger defense. First, A.L.’s story was never crystal clear, suggesting that perhaps a well-informed cross-examination and a contradicting defense witness could have raised a reasonable doubt as to its veracity. Second, all of the people involved in the crime and its revelation were extremely interrelated, suggesting that there is a convoluted back-story that could have provided context for the jurors to understand the accusation and resolve testimonial inconsistencies. Third, Becky Shaffer was the sister-in-law of Vasquez, at the place of the attack on the night it occurred, and with A.L. immediately her accusation, all suggesting that she may have had information useful to the first two facts. Fourth, he had discovery materials from the state that, among other things, contained the EMS report and the names and contact information of the people the police investigated, including Becky Shaffer.
It is against this background that Butler’s decisions against investigation are un*117reasonable. Our recent decision in Brown v. Smith illustrates how the thinness of the prosecution’s case affects our judgment regarding an attorney’s decisions against investigation. 551 F.3d 424 (6th Cir.2008). In that case the petitioner, also convicted of child sexual abuse, alleged that his attorneys were ineffective for failing to interview the psychiatrist who had treated the accuser and lone witness. We explained that because the “entire case hinged on the credibility of [the accuser], and defense counsel were aware that [the psychiatrist] ... had treated her near the time of the alleged assault,” it was “negligent — indeed, constitutionally deficient— for Brown’s attorneys not to seek in camera review of the counseling records.... ” Id. at 431. Similar reasoning applies here: Butler knew that Becky Shaffer was ideally situated to provide information on the only avenue of defense available to Vasquez and that the EMS report contained information contradicting the version of events that the state was alleging at trial. He was therefore deficient by declining to pursue these leads.
Perhaps if Butler knew of other witnesses certain to be cooperative or physical evidence tending to show innocence, it would have been reasonable for him to focus his limited time and resources on developing a different defense based on that evidence. But he did not and so it was not reasonable to interview no potential witnesses and develop no defense. See Williams v. Washington, 59 F.3d 673, 681 (7th Cir.1995) (“Because investigation [of the witnesses] might have revealed evidence bearing upon credibility (which counsel believed was the sole issue in the case), the failure to investigate was not objectively reasonable.”); see also Bigelow v. Williams, 367 F.3d 562, 572 (6th Cir.2004) (holding counsel’s representation deficient because he failed to investigate a possible alibi witness once the witness’s identity was discovered).
The state court held that Butler’s decision was reasonable as to Becky Shaffer because of Kitchen’s statement and because Becky’s husband was a state’s witness. Kitchen’s statement makes sense as a basis to discount Karra Vasquez — Butler had to go through Kitchen to contact her— but less so for Becky Shaffer. Becky could be contacted independently and, while she was a member of the family (Vasquez’s sister-in-law) and may have had reasons not to cooperate, it is incredible for Butler to base a decision against interviewing her on a single statement from the witness’s mother. As to Becky’s husband Don Shaffer being a state’s witness, that could explain a decision not to call Becky but not why he did not attempt to talk with her in the first instance. Cf. Brown, 551 F.3d at 432 (“[0]ur quarrel is not with trial counsels’ decision to forgo calling ... a witness per se, but rather with the lack of any reasonable, timely investigation into what she might have offered the defense.”). Moreover, neither argument takes into account that Butler knew that Becky Shaffer was likely to have relevant information and talking with her could help a develop an otherwise nonexistent defense.
Using the EMS run-sheet to impeach A.L. and the other government witnesses (and as a basis for further investigation) should have been an even more straightforward decision for Butler. Through discovery, Butler had direct knowledge of the EMS report, which contained information demonstrating inconsistencies in A.L.’s story: she described the attack then as “he put it in me” and not as Vasquez licking her vagina, and she claimed that she had taken a shower on the night of the attack (which contradicts her father’s testi*118mony that she fell asleep on the couch with her father).
The state court reasoned that there was no contradiction between the EMS report and the testimony because its possible that (1) the “it” in the EMS description was Vasquez’s tongue and (2) that A.L. showered after her father went to sleep. This post hoc harmonization is not impossible as a matter of logic but does not explain Butler’s indifference to the potential inconsistency (the disharmonious explanation is probably more plausible). He did not try to find out whether or not the testimony of A.L. and her father would stand up to impeachment on the basis of the EMS report. Moreover, Butler should have identified the report as particularly probative because it was made so close to the original report of the attack, so its inconsistency would sharply undercut the corroborative testimony based on interviews with A.L. that took place much later. For instance, penetration and licking are much different acts and the fact that A.L. may have changed her story between the two should have been recognized as premium impeachment evidence. The report further signals an opportunity to track down and interview the EMS respondent, who would have been in a position to testify as to A.L.’s demeanor in the ambulance on the way to her exam, possibly corroborating Becky Shaffer’s testimony that A.L. was at ease in the ambulance and more concerned with the attention she was receiving than disturbed by the abuse she reported.
The warden argues that this line of argument is an impermissible second-guessing of Butler’s strategic decision making. Accepting the patently debatable premise that the omissions discussed in this opinion were the result of a conscious strategy on Butler’s part, the foregoing account of what Butler knew (and what the prospects for Vasquez’s defense were) when he decided against continued investigation demonstrates that decision was “objectively unreasonable because it was a decision made without undertaking a full investigation” and therefore not due deference. Towns, 395 F.3d at 259 (internal quotations omitted); see also Workman v. Tate, 957 F.2d 1339, 1345 (6th Cir.1992) (“Where counsel fails to investigate and interview promising witnesses, and therefore has no reason to believe they would not be valuable in securing defendant’s release, counsel’s inaction constitutes negligence, not trial strategy.”) (internal quotations omitted). To this, we add only that the pretrial investigation actually admitted to by Butler — his legal research on sex abuse and child rape cases — should have pointed him toward the investigation he failed to undertake. It is well known in the literature (and the cases cited above) that the credibility of the child witness is often central to the success of child sex abuse prosecutions and that the circumstances surrounding the initial accusation of the abuse are important indicia of credibility. See generally, e.g., Kamela London et al., Disclosure of Child Sexual Abuse: What Does Research Tell Us About the Ways that Children Tell?, 11 PSYCH. PUB. POL. & L. 194 (2005) (discussing various psychological models of children’s accusations of child sexual abuse and their importance to understanding the accuracy of testimony in child sexual abuse trials). This suggests that, even without knowing what Butler knew about the available evidence, Becky Shaffer and the EMS run-sheet would be natural starting points to an investigation and a defense. Yet, Vasquez’s attorney did not pursue these leads. We do not believe that this behavior can accurately be described as a “strategic choice”. Instead, we hold it to be representation “below an objective standard of reasonableness” and constitutionally defi-*119dent. Strickland, 466 U.S. at 688, 104 S.Ct 2052.
B
This deficient representation only amounts to a violation of Vasquez’s Sixth Amendment rights if it resulted in prejudice — a probability of a different result that is “sufficient to undermine confidence in the outcome.” Strickland, 466 U.S. at 694, 104 S.Ct. 2052. In determining whether Vasquez has carried his burden, we must evaluate the deficiency in light of the “totality of the evidence before the ... jury.” Id. at 695, 104 S.Ct. 2052. And so, in a case in which the “verdict or conclusion [is] only weakly supported by the record is more likely to have been affected by errors....” Id. at 696, 104 S.Ct. 2052. Accordingly, in cases like this one, where the only evidence of the crime or the defendant’s guilt is the testimony of the victim, our circuit has been especially willing to find prejudice from deficient representation because “[t]he lack of physical evidence confirming sexual activity meant that this was necessarily a close case at the trial level.” Hodge v. Hurley, 426 F.3d 368, 386 (6th Cir.2005).
The missing evidence here undermines our confidence that had Butler pursued the leads he had immediately available to him the same outcome would have obtained. The EMS report demonstrates a specific and non-trivial difference in her story over how the attack took place: the run-sheet implies A.L. reported that there was penetration by Vasquez’s penis in contradiction to the evidence at trial.3 Becky Shaffer’s account reveals inconsistences in A.L.’s timeline; that S.L. coached A.L. and was prompting responses to the police; that she acted relaxed and not in conformity with a girl recently revealing a traumatic experience; and specific revelations that A.L.’s concern was over attention (would she get on TV) and not the crime. Speaking with her would have also led Butler to additional defense witnesses. The testimony from Salopek and her daughter, in turn, would have provided detailed reasons that A.L. is not trustworthy and does not understand the consequences of her lies; that she was not acting out-of-the-ordinary the week after she revealed the attack; and that A.L. had just learned of the sexual abuse suffered by Snyder. Together, this evidence challenges the specifics of her story, her motivations, and her general truthfulness. Moreover and equally important, it undermines the corroboration evidence presented by the state — that A.L.’s story did not change in her retelling of it, that she acted out of character after the attack, and the implausibility of a nine-year old making up such a sexual-molestation story.
The warden makes two arguments against the prejudicial effect of this missing evidence. First, the warden argues that Becky Shaffer may not have in fact *120cooperated because she declined to give a written statement to the investigating police officer. This fact is not as significant as the warden would have it. Shaffer did speak with the officer, but chose not to give a written statement after her husband had already given one. It is true that she did not, on her own volition, volunteer facts to the police officer, but that is not evidence of what she would have said had she been directly asked. Indeed, legally trained defense counsel are likely to ask different questions and bring a different context to the investigation than the investigating police officer. The police officer was trying to establish the facts of the attack. Butler, undertaking his investigation with an understanding of the state’s case, would be building a case specifically against A.L.’s version of events. He likely would have asked different questions and placed greater weight on information to present a jury. For instance, he may have focused on the ambulance ride (because it would be the best evidence of how A.L. acted immediately after accusing Vasquez) and on information about A.L.’s credibility as a witness, rather than on what happened the night of the attack. According to Becky Shaffer’s testimony, such questions would have elicited the answers that now provide important impeachment evidence.
Second, the warden argues that the additional evidence was cumulative on Butler’s attacks against A.L.’s credibility. To be sure, “evidence that is merely cumulative of that already presented does not ... establish prejudice.” Getsy v. Mitchell, 495 F.3d 295, 313 (6th Cir.2007) (en banc) (internal quotations omitted). Our cases, however, do not tell us clearly when evidence becomes sufficiently different to no longer be “cumulative” or at what level of generality one must compare the evidence. In our most skeptical formulation, we have said, “in order to establish prejudice, the new evidence that a habeas petitioner presents must differ in a substantial way— in strength and subject matter — from the evidence actually presented.... ” Hill v. Mitchell, 400 F.3d 308, 319 (6th Cir.2005). In light of the weakness of the defense at trial, the missing evidence here satisfies even this high standard.
Butler used the cross-examination of the prosecution’s witnesses to attack A.L.’s credibility and to imply that her story changed. For instance, A.L. admitted that she sometimes lies and that her father has to punish her for that. A.L.’s father agreed that he sometimes had to punish her for lying. He admitted that he repeatedly went over the accusation with her to make sure she was telling the truth. He also admitted that during the eleven days between the incident and her report of it, he did not hear anything to indicate A.L. (who was visiting relatives) acted out of the ordinary. Further, Butler used A.L.’s father to demonstrate differences between the version of the attack in the police report and the version recorded as the attack history in the hospital, where A.L. did not tell the doctor that Vasquez covered her mouth or that the attack ended when her father called for her. Similarly, on cross examination Don Shaffer indicated that those same details in the hospital report were not in A.L.’s initial description of the attack to him. Butler relied on this testimony to create a narrative in closing about A.L.’s credibility, suggesting that she was caught in a lie by her father, but insisted on the truth of the accusation to avoid the consequences of having lied about it to her father.
The missing evidence is not cumulative on these narrow admissions that Butler induced on cross-examination. The missing evidence is not a restatement of the general proposition of A.L.’s untrustwor-thiness or the inconsistencies already *121raised. Cf. Hartman v. Bagley, 492 F.3d 347, 361 (6th Cir.2007) (holding no prejudice where petitioner’s main argument is only “contentions that counsel failed to expand on or corroborate or fully develop the factors listed in” his psychologist’s report that were testified to in front of the jury). Instead, the evidence provides direct and specific challenges to her motivations (she wants to be on TV, she wants attention), her trustworthiness (she lies a lot and her friends and friends’ parents confirm that she does not understand consequences of her lies), the stability of her story (the change in the EMS report, the prompting from her father), and the origins of her story (she just recently learned Ashley Snyder had been molested), each of which was not presented to the jury via Butler’s cross-examinations. The inconsistencies elicited between the hospital report and her initial report also were only sins of omission, unlike the EMS report, which contains information possibly contradicting the central fact of the crime.
Moreover, Butler’s attempts at impeachment all relied on the testimony of the state’s witnesses and frequently were likely unconvincing because they relied only on the state’s witnesses. For instance, A.L.’s admission that she “sometimes lies” is less impressive without corroboration from additional witnesses that underscore that she lies specifically to get attention and that she frequently does not understand the consequences of her lies. Cf. Clinkscale, 375 F.3d at 445 (holding that corroboration evidence to defendant’s alleged alibi was not cumulative because without “any corroborating witnesses [the defendant was] left ... without any effective defense.”) (internal quotations omitted). Butler’s other attempts at impeachment were followed by flat rejections of the broader propositions that he tried to establish. S.L. rejected the proposition that his daughter “had a tendency to make things up.” Similarly, Don Shaffer denied on cross that A.L. frequently made things up beyond “ordinary, kids fighting.” The missing evidence, on the other hand, was not subject to immediate denial and so would have been less vulnerable to being discounted by the jury. See Bigelow, 367 F.3d at 565 (holding that additional alibi witnesses were not cumulative because of the vulnerability of the one alibi witness called).
Where, as here, the evidence that the jury did not hear provides specific impeachment evidence, we have not held it to be “cumulative” merely because other, less convincing evidence that also happened to be “impeachment evidence” was heard. See, e.g., Brown, 551 F.3d at 434-35. And this outcome makes sense, even under the “substantial difference” standard in Hill: the ultimate question is not the analytical difference between the actual trial and a hypothetical one, but whether the fact of missing evidence “undermine[s] confidence in the outcome.” Strickland, 466 U.S. at 694, 104 S.Ct. 2052.
Accordingly, we hold that, in light of centrality of credibility to the prosecution’s case and the strength of the missing impeachment evidence, Vasquez has demonstrated that but for his counsel’s unprofessional errors there was a reasonable probability that his trial would have had a different outcome; our confidence in the verdict is indeed undermined.
IV
For the reasons explained, Robert Vasquez received constitutionally ineffective assistance of counsel. The Ohio courts’ holding decision of no prejudice was contrary to clearly established Supreme Court precedent and Vasquez is therefore entitled to the writ of habeas corpus. The *122judgment of the district court is AFFIRMED.
. The Ohio appellate courts review a trial court's decision as to a post-conviction petition for relief under an abuse of discretion standard. See, e.g., State v. Gondor, 112 Ohio St.3d 377, 860 N.E.2d 77, 86 (2006). This raises an apparently unresolved question about the state decision a federal court should defer to under AEDPA. Where an appellate court reviews under an "abuse of discretion” standard, it is not adjudicating the merits of the claim so much as the merits of the decision under review. The question is only whether it fell into what is presumably a broad band of discretion. AEDPA, however, is meant to secure deference to the application of federal law by a state court in its "adjudicat[ion] on the merits" of a petitioner's claim. 28 U.S.C. § 2254(d). Accordingly, AEDPA perhaps does not have anything to say about a state court's decision that another state court did not abuse its discretion — it is concerned only with whether the actual adjudication of the merits of petitioner’s claim was contrary to, or an unreasonable application of, federal law. On this view, we should focus our analysis on the lower court's adjudication and not the appellate court. This approach finds analogical support in our court's refusal to give AEDPA deference to a state appellate court review for plain error. See Benge v. Johnson, 474 F.3d 236, 246 (6th Cir.2007) ("Because Benge could have met his burden under Strickland despite not being able to demonstrate plain error, this analysis did not constitute an 'adjudication on the merits’ of Benge’s ineffective-assistance-of-counsel claim.”) (quoting 28 U.S.C. § 2254(d)). Other cases, however, have focused on the reasoning actually followed by the state court and not the standard of review applied. See Fleming v. Metrish, 556 F.3d 520, 530-32 (6th Cir.2009) (distinguishing Benge and holding a review for plain error is an adjudication on the merits where the state appellate court first determined the merits of the claimed error before holding that it did not effect substantial rights).
Nevertheless, we do not find it necessary to resolve this issue here because we hold that the appellate court also applied law contrary to the Strickland standard by asking whether the outcome would have been different. We note that it is plausible that a state appellate court might, without repeating the mistake of a lower court, still affirm because it is applying a deferential standard of review.
. The warden attributes this failure to Vasquez. It is true that Vasquez only told Butler to contact his wife and mother-in-law and, apparently, did not mention Becky Shaffer or anyone else as a potential witness. This might be a persuasive rejoinder to Vasquez’s allegations if Becky Shaffer were not such an obvious person to contact because: (1) she is Vasquez’s sister-in-law (at the very least, Butler may have contacted her to get in touch with Karra Vasquez); (2) the crime took place in her home; (3) she was present in the home at the time it occurred; and (4) she was present immediately after A.L.'s accusation.
This attribution is part of a strategy, on the part of both the warden and the state court, to place the attorney's failure to investigate at Vasquez's feet. While we are sympathetic to the defense attorney who has a client that will not help, we do not believe our cases to support the proposition that a defendant gets only the defense diat he is capable of providing personally. (They hold only that it is not unreasonable for an attorney to limit investigation once her client has frustrated her otherwise reasonable efforts. See, e.g., Lorraine v. Coyle, 291 F.3d 416, 435 (6th Cir.2002)). The Sixth Amendment guarantees minimally effective representation because the adversarial testing of the state’s case, a cornerstone of our criminal justice system, is very difficult without counsel. See Kimmelman v. Morrison, 477 U.S. 365, 384, 106 S.Ct. 2574, 91 L.Ed.2d 305 (1986) (“In making the competency determination, the court should keep in mind that counsel’s function ... is to make the adversarial testing process work ....") (internal quotation marks omitted). An attorney (especially an experienced one like Don Butler) has skills and knowledge beyond the ken of an average criminal defendant. He should be expected to take a sober account of the case he is presented with and proceed to put together the best challenge to the prosecution’s proof. This is true even if he must use information gleaned from discovery and investigation instead of the defendant’s mouth.
. The state court’s harmonization that the "it” could be Vasquez’s tongue is not responsive to our focus about how the jury would likely have taken it. Coupled with the inconsistency about the shower, we think it likely a jury would consider the run-sheet evidence of instability in A.L.’s story. This is especially so because it is unlikely that the jury would believe a nine-year-old would colloquially describe Vasquez's licking as ”put[ting] it in,” whereas she might be likely to use that term for penetration. Moreover, unlike A.L.'s story at trial, penetration is not consistent with the circumstances of the attack occurring in a short window of time, with her father’s presence upstairs, and on the same bunk-bed where Vasquez's children were sleeping. In any case, the fact that we cannot be sure about how the jury would take the evidence does not prevent a finding of prejudice, since Vasquez need only undermine our confidence in the outcome, not prove that a different outcome was certain with competent counsel.