*578 OPINION

BOGGS, Chief Judge.

This case presents the question of what type of deference we owe on collateral review to a state court’s harmless-error determination. David Eddleman was convicted of second-degree murder and a firearm offense in a Michigan state court. On direct review, the Michigan Court of Appeals affirmed his conviction, concluding that the trial court erred in admitting his coerced confession but that the error was harmless. Eddleman petitioned for a writ of habeas corpus in federal court. The district court granted the writ, and warden Ken McKee appealed.

We affirm. We hold that, when a state court has found an error to be harmless, we should ask on collateral review whether the state court’s harmless-error decision was contrary to, or an unreasonable application of, the clearly established federal rule that a trial error is harmless only if it is harmless beyond a reasonable doubt. Applying this standard of review to the case at hand, we hold that the Michigan Court of Appeals’s harmless-error determination was an unreasonable application of the Supreme Court’s decisions in Chapman v. California, 386 U.S. 18, 87 S.Ct. 824, 17 L.Ed.2d 705 (1967), and Arizona v. Fulminante, 499 U.S. 279, 111 S.Ct. 1246, 113 L.Ed.2d 302 (1991).

Eddleman’s conviction arose out of the shooting death of sixteen-year-old Joane Georgescu early in the morning of October 13, 1996. Georgescu was seated in a car parked near the corner of Kirkland and Trenton Streets in western Detroit, near Dearborn, when a bullet shot from a passing car entered through the trunk of her car, passed through the back seat, and struck her in the heart. The state of Michigan charged Eddleman with first-degree premeditated murder, second-degree murder, and possession of a firearm during a felony. He stood trial in Wayne County Circuit Court, starting on July 15, 1999.

The government argued that Eddleman was a member of the Insane Spanish Cobras, a street gang, and shot Georgescu while carrying out a mission ordered by the gang’s leader, Jesus Garcia. The evidence it introduced to support its theory consisted of Eddleman’s confession and the testimony of four key witnesses — three other gang members and a jailhouse informant. As discussed below, reasons existed to doubt the credibility of each of the witnesses. The government did not present any physical evidence linking Eddle-man to the crime or any other eyewitnesses who could identify the vehicle used in the shooting or any of its occupants.

The first witness, Brian Babbitt, testified that he was a member of the Insane Spanish Cobras street gang in October 1996 and that Garcia was the gang’s leader. He stated that on the evening of October 12, 1996, shortly before the shooting, Garcia ordered Eddleman on a mission to find and shoot members of a rival gang in order to increase his rank within the Insane Spanish Cobras. According to Babbitt, Eddleman received a military-style M-l carbine, the type of gun used in the shooting, and then left to carry out the mission.

Eddleman allegedly sat in the front passenger seat of a burgundy Chevrolet Nova carrying three other people: fellow gang member Richard Glockens, who was driving; Babbitt, who was in the back seat; and an unnamed fourth individual, who *579 was lying across the back seat after passing out due to alcohol consumption. Their initial attempt to find rival gang members failed. They returned to the gang house, where, according to Babbitt, Eddleman told Garcia that he knew the location of a rival gang party. Garcia told him to “shoot it up,” and the car departed again, with Eddleman giving Glockens directions to an address near the intersection of Kirkland and Trenton Streets. A second car followed, “[t]o make sure the mission was done right.” Babbitt claimed that, before the shooting, he got out of Eddle-man’s car and got into the second car, because he “had a bad feeling” about what was happening. From that second car, he testified, he saw a total of ten to twelve shots fired from the passenger’s side of the Nova toward a crowd of young men gathered outside a house. According to Babbitt, “Dave Eddleman [was] the only one firing any shots.”

Two categories of facts brought out at trial call Babbitt’s credibility into question. First, Babbitt received significant benefits from the police in exchange for his testimony. On January 3, 1997, Babbitt was arrested for the killing of Georgescu. On January 7, 1997, he was arrested again, this time for the murder of Freddy Sanchez. The Wayne County Prosecutor’s office granted him immunity from prosecution for both murders on January 22, 1997, in exchange for his testimony against Ed-dleman, Garcia, Glockens, and another gang member allegedly involved in the Sanchez murder. At the time, Babbitt also faced charges of felony assault, using a firearm in the commission of a felony, and violating probation. After negotiating the immunity agreement, the prosecutor allowed Babbitt to plead guilty only to a reduced charge of aiming without malice, with a sentence of time served.

Second, Babbitt’s trial testimony conflicted in three ways with statements he had made previously. On January 3, 1997, and again on January 10, 1997, Babbitt told police that he was never in a car with Eddleman on the night of the shooting; at trial, he testified that he was in the same car as Eddleman until shortly before the shooting. In the. same January 3 interview, Babbitt stated that Eddleman brought the M-l carbine used in the shooting himself; in his preliminary examination testimony for a separate trial against Garcia, he claimed that Garcia had given Eddleman the gun; finally, at trial, he testified that another gang member, Cano, gave Eddleman the gun. In another interview with police on January 7, 1997, Babbitt did not mention that Garcia was involved in the shooting; at trial, he testified that Garcia directly ordered the shooting.

The next witness, jailhouse informant Ricky O’Neal, testified that he was the head of a Chicago-area gang, that he had met and befriended Garcia while in prison in 1993 or 1994, and that Garcia had introduced him to Eddleman while Garcia and Eddleman were in jail awaiting trial for Georgescu’s murder. O’Neal claims that he talked to Eddleman while the two of them, along with Garcia and Glockens, were working out at the jail gym. According to O’Neal, Eddleman admitted that he had “shot and killed that innocent girl” and stated that “he wanted to take the rap alone to free” Garcia and Glockens.

O’Neal testified pursuant to a plea agreement in which the state agreed to reduce a pending charge of assault with intent to rob, which carried a possible life sentence, to two counts of felony assault, for which he received concurrent sentences of one to four years. He had an extensive prior criminal record, including convictions *580 for breaking and entering, larceny, and unarmed robbery.

The next two witnesses, Brian Weaver and Thomas Valastek, both testified that they heard Eddleman admit to the shooting. Weaver testified that he attended a gang meeting on October 13, 1996, the day after Georgescu’s murder, and that Eddleman stated at that meeting that he had committed a drive-by shooting while on a mission the previous night. Weaver explained that he remembered the date because, on the previous evening, he had watched the boxing match between Mike Tyson and Evander Holyfield in which Tyson bit Holyfield’s ear. On cross-examination, the defense impeached Weaver’s testimony with evidence that Tyson and Holyfield did not fight each other at all until November 9, 1996, and that the ear-biting bout did not take place until June 28, 1997. The defense also presented evidence that Weaver may have avoided punishment for numerous probation violations by testifying against Eddleman.

Valastek gave testimony substantially similar to Weaver’s: he stated that he was watching the Tyson-Holyfield boxing match, that he heard Garcia mention a mission he wanted completed that evening, and that the next morning he heard Eddle-man bragging that he had been involved in a drive-by shooting the day before. In addition to the inaccuracy of the boxing reference, three other pieces of information call into question Valastek’s credibility. First, he too cooperated with police only after he was arrested in connection with Georgescu’s murder. Second, he admitted on cross-examination that his initial statement to police contained “a lot of lying.” Third, he admitted that he chose to cooperate fully with police a day after he saw that Babbitt was receiving special perks in jail — such as family visits, outside food, and television privileges — and figured “he must have told them something really good to give him all that.”

A fifth witness, Jennifer Lertola, Eddle-man’s girlfriend at the time of the shooting, testified that Eddleman called her shortly after the time of the shooting and told her to watch the Channel 7 news. Before trial, she allegedly had told police that, when she asked Eddleman if he had shot Georgescu, Eddleman laughed and said “that’s for me to know and for you to find out,” and that Eddleman later admitted to the shooting. At trial, however, she stated that she did not think that she made those statements and that any statements she did make were driven by fear that the police would take her children away if they thought she was concealing information.

Eddleman argued that the Insane Spanish Cobras had no role in the shooting. Instead, he claimed at trial, the person who shot Georgescu was targeting a resident of a house at the corner of Kirkland and Trenton Streets, who had been involved in an earlier bar fight. Four witnesses supported this core theory. On cross-examination, all four admitted that they formed their beliefs about what happened based on rumor and speculation rather than specific evidence.

The prosecution made frequent references to Eddleman’s confession, both in supporting its theory of the case and in rebutting Eddleman’s theory. Four references stand out. First, in his opening statement, the prosecutor emphasized the confession as soon as he began to discuss the facts:

Let me see, if I can, introduce you once again to the Defendant in this case.... This is David Eddleman, Insane Spanish Cobra, a member of the Insane Spanish Cobras. Back in October of 1996, the evidence will show you that this Defendant, this Insane Spanish Cobra, David Eddleman, confessed, made a full con *581 fession to shooting and killing an innocent fifteen-year-old [sic] girl....

Second, Detective Barbara Simon read Eddleman’s confession to the jury. Her testimony, focused entirely on the confession, lasted for almost five hours. Third, the prosecutor used the confession as a potent weapon during his cross-examination of each of the four witnesses supporting Eddleman’s “prior fight” theory. 1 Fourth, the prosecutor began his summary of the evidence in his closing argument by stating that “[w]e’ve got the Defendant’s confession, first and foremost of all in this case,” and later referred to the confession at least five more times.

The jury struggled to reach a verdict. After a day of deliberations, the jury communicated to the court that it was deadlocked. The court instructed the jurors to continue their deliberations. About an hour later, the jury returned a second communication stating that it could no longer deliberate and had reached an impasse. The court then dismissed the jury for the weekend. Shortly after reconvening, the jury found Eddleman not guilty of first-degree murder but guilty of both second-degree murder and the firearm charge. The trial court sentenced Eddle-man to thirty to sixty years of imprisonment for the murder conviction, to be served consecutively to two years of imprisonment for the firearm conviction.

Eddleman appealed to the Michigan Court of Appeals, arguing, inter alia, that the trial court erred in admitting the confession into evidence. The court agreed, but concluded that the error was harmless:

Nonetheless, we conclude that reversal is not warranted because the error was harmless. People v. Anderson, 446 Mich. 392, 406, 521 N.W.2d 538 (1994). The testimony of several witnesses clearly implicated the defendant in the crime. Brian Babbitt, also a member of defendant’s street gang, was in a car traveling behind the car defendant was riding in on the night of the shooting. Babbitt testified that he witnessed shots coming from the passenger side of defendant’s car, where defendant was seated. Two other gang members, Brian Weaver and Thomas Valastek, testified that they heard defendant admit to being the shooter. Defendant also told a fellow inmate that he had shot and killed Georgescu. Given the weight of this evidence, we conclude that the erroneous admission of the confession was harmless beyond a reasonable doubt. Id.

People v. Eddleman, No. 224957, slip op. at 2, 2002 WL 433338 (Mich.Ct.App. Mar.19, 2002). The Michigan Supreme Court denied his petition for review.

Eddleman then filed a timely federal petition for a writ of habeas corpus pursuant to 28 U.S.C. § 2254, attacking only the state appellate court’s harmless-error determination. The district court conditionally granted the writ, mandating that the state either schedule a new trial within ninety days or release Eddleman. The warden appealed, and the district court agreed to stay its order pending the resolution of this appeal.

The admission of an involuntary confession at trial is subject to harmless-error analysis. Arizona v. Fulminante, 499 U.S. 279, 303, 111 S.Ct. 1246, 113 *582 L.Ed.2d 302 (1991). In this case, we must clarify what type of deference we owe on collateral review to a state court’s determination that an error was harmless.

Prior to the enactment of the Antiter-rorism and Effective Death Penalty Act of 1996 (“AEDPA”), the Supreme Court articulated two harmless-error standards. On direct review, “before a federal constitutional error can be held harmless, the court must be able to declare a belief that it was harmless beyond a reasonable doubt.” Chapman v. California, 386 U.S. 18, 24, 87 S.Ct. 824, 17 L.Ed.2d 705 (1967). On collateral review, however, the Court instructed us to find an error to be harmless unless the error “had a substantial and injurious effect or influence in determining the jury’s verdict.” Brecht v. Abrahamson, 507 U.S. 619, 637, 113 S.Ct. 1710, 123 L.Ed.2d 353 (1993) (quoting Kotteakos v. United States, 328 U.S. 750, 776, 66 S.Ct. 1239, 90 L.Ed. 1557 (1946)).

When Congress enacted AEDPA, it complicated this dichotomy. AEDPA states:

An application for writ of habeas corpus on behalf of a person in custody pursuant to the judgment of a State court shall not be granted with respect to any claim that was adjudicated on the merits in State court proceedings unless the adjudication of the claim—

(1) resulted in a decision that was contrary to, or an 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). When a state court addresses a harmless-error claim on direct review, the clearly established federal law that it must apply is the Chapman standard. As such, AEDPA’s plain language requires federal courts to deny a habeas petition in this situation unless a state court’s determination of harmless error was contrary to, or an unreasonable application of, the principle that an error must be harmless beyond a reasonable doubt.

In cases when a state court ruled that an error was harmless, however, it is not clear whether this new standard — Chap man plus AEDPA deference — supersedes Brecht’s “substantial and injurious effect” standard or supplements it. The Second, Seventh, and Eighth Circuits have held that AEDPA supersedes the Brecht standard and therefore have applied only Chapman’s “harmless beyond a reasonable doubt” standard plus AEDPA deference. See Ben-Yisrayl v. Davis, 431 F.3d 1043, 1052 (7th Cir.2005); Gutierrez v. McGinnis, 389 F.3d 300, 306 (2d Cir.2004); Whitmore v. Kemna, 213 F.3d 431, 433 (8th Cir.2000). The Fourth and Ninth Circuits have held that AEDPA supplements Brecht and therefore have crafted a two-step approach, granting habeas relief only if the state court’s decision was contrary to, or an unreasonable application of, the Chapman standard and the error is not harmless under the Brecht standard. See Inthavong v. Lamarque, 420 F.3d 1055, 1059 (9th Cir.2005); Jones v. Polk, 401 F.3d 257, 264-66 (4th Cir.2005). Joined by the First, Fifth, and Tenth Circuits, see Robertson v. Cain, 324 F.3d 297, 306-07 (5th Cir.2003); Herrera v. Lemaster, 301 F.3d 1192, 1200 (10th Cir.2002) (en banc); Medina v. Matesanz, 298 F.3d 98, 101 (1st Cir.2002), we have agreed that AEDPA supplements Brecht, but concluded that, if the petitioner can make the showing required by Brecht, “he will surely have demonstrated that the state court’s finding that the error was harmless beyond a reasonable doubt — the Chapman standard— was outside the realm of credible out *583 comes, and therefore resulted from an unreasonable application of Chapman.” Nevers v. Killinger, 169 F.3d 352, 371-72 (6th Cir.1999). Consequently, we have continued to apply only the Brecht standard despite the enactment of AEDPA. See Bulls v. Jones, 274 F.3d 329, 335 (6th Cir.2001).

Today, we reconsider our position in light of the Supreme Court’s decision in Mitchell v. Esparza, which strongly implied that courts should apply only the Chapman plus AEDPA deference standard of review. 540 U.S. 12, 17-18, 124 S.Ct. 7, 157 L.Ed.2d 263 (2003) (per curiam), rev ‘g 310 F.3d 414 (6th Cir.2002). In Esparza, the Court reversed this court’s grant of a writ of habeas corpus, holding that we showed insufficient deference under AEDPA to the harmless-error analysis of the Ohio Court of Appeals and explaining:

A constitutional error is harmless when “it appears ‘beyond a reasonable doubt that the error complained of did not contribute to the verdict obtained.’ ” Neder [v. United States, 527 U.S. 1, 15, 119 S.Ct. 1827, 144 L.Ed.2d 35 (1999) (quoting Chapman, 386 U.S. at 24, 87 S.Ct. 824).] We may not grant respondent’s habeas petition, however, if the state court simply erred in concluding that the State’s errors were harmless; rather, habeas relief is appropriate only if the Ohio Court of Appeals applied harmless-error review in an “objectively unreasonable” manner.

Id. at 17-18, 124 S.Ct. 7 (some citations omitted). The Supreme Court never discussed or cited Brecht. Two circuits, the Second and the Seventh, already have relied on Esparza to revisit their prior precedents and adopt the standard of Chapman plus AEDPA deference. See Ben-Yisrayl, 431 F.3d at 1051 (overruling sub silentio Aleman v. Sternes, 320 F.3d 687, 690-91 (7th Cir.2003), which applied the two-step approach); Gutierrez, 389 F.3d at 306. The Fifth Circuit has suggested in dicta that it might do the same. See Garcia v. Quarterman, 454 F.3d 441, 447 (5th Cir.2006). Taking guidance from Esparza, we now join those circuits. 2 We hold that AEDPA replaced the Brecht standard with the standard of Chapman plus AEDPA deference when, as here, a state court made a harmless-error determination. 3

*584 This result makes practical sense for two reasons. First, a judicially crafted standard generally gives way when Congress steps in to address the same issue, as long as the original standard was not constitutionally required. See, e.g., Schweiker v. Chilicky, 487 U.S. 412, 425, 108 S.Ct. 2460, 101 L.Ed.2d 370 (1988) (holding that the court-created Bivens remedy is unavailable when Congress, through the Social Security Act, provided an alternative mechanism for a plaintiff to vindicate his due process rights). That is precisely what happened here. The Brecht Court saw the harm caused by de novo collateral review of state court harmless-error determinations and crafted its own solution to prevent that harm. See 507 U.S. at 636, 113 S.Ct. 1710. It recognized that in mandating an alternative standard of review it was acting as a temporary substitute for Congress, not as a permanent partner: “In the absence of any express statutory guidance from Congress,” it explained, “it remains for this Court to determine what harmless-error standard applies” to the petitioner’s claim. Id. at 633, 113 S.Ct. 1710 (emphasis added). When Congress enacted AEDPA it obviated the need for such a stopgap.

Second, we owe a different type of deference to state proceedings when a state court has found an error to be harmless than when it found no error. We defer to state-court judgments for two reasons. First, we respect the “presumption of finality” that “attaches to the conviction and sentence at the conclusion of direct review.” Brecht, 507 U.S. at 633, 113 S.Ct. 1710. Second, we respect a state’s interest, grounded in principles of comity, in maintaining the “primary authority for defining and enforcing the criminal law.” Id. at 635, 113 S.Ct. 1710. Both Brecht and AEDPA reflect these finality and comity interests. See, e.g., 141 Cong. Rec. H1425, 1427 (daily ed. February 8, 1995) (statement of Rep. Cox) (“The central problem underlying federal habeas corpus review” prior to AEDPA “is a lack of comity and respect for state judicial decisions.”); 141 Cong. Rec. H1400, 1400 (daily ed. February 8, 1995) (statement of Rep. McCollum) (AEDPA “stands for the clear and simple proposition that there must be finality and accountability.”).

One might think that these finality and comity interests are strongest when a state court addressed the harmless-error issue and concluded that an error was harmless. In one sense, that intuition is correct. When a state court explicitly has held that an error was harmless, we owe deference to that conclusion of law as a matter of comity. When a state court does not reach the question of harmless error, we owe no such deference, since there is no relevant conclusion of law to which we could defer.

However, we do not defer to a state court solely out of respect for its resolution of a particular legal issue. The principles of comity and finality also dictate that we should defer to a state court’s ultimate disposition of a case. For two reasons, this second type of deference is weaker when a state court found an error to be harmless than when it found no error at all. First, the state interest in finality is weaker. A state court acknowledges that its resolution of the case was not perfect *585 when it recognizes that a federal constitutional error occurred at trial. While an imperfect judgment still is entitled to a presumption of correctness, this presumption is not as strong as it would be absent any finding of error. Second, the relevant state interest in comity is weaker. When a state proceeding has violated federal constitutional rights, the state’s interest in the autonomous administration of its criminal law gives way, at least in part, to the federal government’s interest in protecting those rights. When a state court itself admits that a constitutional error occurred at a state trial (and therefore that a strong federal interest is in play), we do not compromise the state’s status as a separate sovereign by lessening the deference to the final outcome reached in state court.

In the situation we face today, in which a state court found error but concluded that the error was harmless, we owe deference primarily to the state court’s harmless-error determination rather than to the final outcome the state court reached. The standard we announce, Chapman plus AEDPA deference, provides this type of deference by requiring us to decide whether that determination was contrary to, or an unreasonable application of, clearly established federal law. In contrast, if the state court had found no error, we would owe deference only to the final outcome the state court reached. The Brecht standard, which clearly applies in that situation, provides that type of deference by requiring us to determine whether an error had a “substantial and injurious effect on the jury’s verdict.”

Thus, our analysis of principles of separation of powers and of federalism supports our reading of the Supreme Court’s precedent. When a state court addressed the question of harmless error, the appropriate standard of review is the Chapman standard plus AEDPA deference.

ill

We now must determine whether the Michigan Court of Appeals decision that admitting Eddleman’s confession was harmless error was contrary to, or an unreasonable application of, Chapman. Despite some confusion about the relationship between AEDPA and Brecht, the district court properly applied the Chapman standard plus AEDPA deference and concluded that the state court’s decision was an unreasonable application of Chapman. We agree.

Because the Michigan Court of Appeals applied the correct legal standard, asking whether the error was harmless beyond a reasonable doubt, we evaluate its decision under the “unreasonable application” portion of AEDPA. See (Terry) Williams v. Taylor, 529 U.S. 362, 405, 120 S.Ct. 1495, 146 L.Ed.2d 389 (2000). A state court unreasonably applies clearly established Supreme Court precedent when it reaches a result that “falls outside the realm of plausible credible outcomes” dictated by the governing law. Barker v. Yukins, 199 F.3d 867, 872 (6th Cir.1999). Since the district court based its factual findings solely “on a transcript from the petitioner’s state court trial,” we review de novo both those factual findings, Wolfe v. Brigano, 232 F.3d 499, 501 (6th Cir.2001), and the district court’s legal conclusions, Greer v. Mitchell, 264 F.3d 663, 671 (6th Cir.2001).

In Arizona v. Fulminante, the Supreme Court applied Chapman’s “beyond a reasonable doubt” harmless-error analysis in the context of a coerced confession. 499 U.S. 279, 295, 111 S.Ct. 1246, 113 L.Ed.2d 302 (1991) (citing Chapman, 386 U.S. at 24, 87 S.Ct. 824). Fulminante represents the best available explication of the “clearly established Federal law, as *586 determined by the Supreme Court of the United States,” relevant to Eddleman’s claim.

The defendant in Fulminante twice confessed to killing his stepdaughter. Id. at 283, 111 S.Ct. 1246. The Court held that the first confession, made to his FBI-informant cell mate while in jail on another charge, was coerced. Ibid. The key piece of evidence remaining against him was his second confession, made to the cellmate’s wife after his release from prison. Id. at 298, 111 S.Ct. 1246. The Court questioned the credibility of testimony about this second confession, noting that the cellmate did not remember that it occurred until a year later, even though he was in the car when the alleged conversation between his wife and the defendant took place; that the wife’s account of the circumstances surrounding the confession sounded somewhat implausible; and that prosecutors agreed to secure the cellmate a sentence reduction and to place both him and his wife in the witness protection program in exchange for their testimony. Id. at 300, 111 S.Ct. 1246.

Because the Court considered the case on direct review, it applied the Chapman standard, id. at 296, 111 S.Ct. 1246, and concluded that the erroneous admission of the first confession was not harmless, id. at 302, 111 S.Ct. 1246. It explained that the erroneous admission of a coerced confession presents a unique problem under Chapman:

A confession is like no other evidence. Indeed, the defendant’s own confession is probably the most probative and damaging evidence that can be admitted against him.... In the case of a coerced confession such as that [at issue in this case], the risk that the confession is unreliable, coupled with the profound impact that the confession has upon the jury, requires a reviewing court to exercise extreme caution before determining that the admission of the confession at trial was harmless.

Id. at 296, 111 S.Ct. 1246 (quotation marks and citations omitted). It then evaluated the impact on Fulminante’s trial of admitting his first confession, focusing on some of the five indicia of harmlessness that it previously had set forth: the overall strength of the government’s case; the emphasis placed on the confession within the government’s case; the relationship between the confession and other evidence; the evidentiary value of the confession; and the defendant’s opportunity to attack the confession through cross-examination. Id. at 296-302, 111 S.Ct. 1246; see Delaware v. Van Arsdall, 475 U.S. 673, 684, 106 S.Ct. 1431, 89 L.Ed.2d 674 (1986). 4

Considering these five indicia of reasonableness, we see a strong congruence between Eddleman’s case and Fulminante. Regarding the strength of the government’s case, the prosecutions in Eddle-man’s case and in Fulminante both rested solely on the testimony of questionably credible witnesses. As in Fulminante, no physical evidence linked Eddleman to the crime. As in Fulminante, the key prosecution witnesses against Eddleman told inconsistent or unbelievable stories: Babbitt’s trial testimony differed in several key ways from the first account he gave to *587 police, while Weaver and Valastek both identified the date of the murder and Ed-dleman’s admitting to it by reference to a boxing match that did not occur until nine months later. And as in Fulminante, those witnesses all had reasons to implicate Eddleman other than a commitment to telling the truth: Babbitt and Valastek cooperated only after they were arrested for Georgescu’s murder; Babbitt, Valas-tek, and O’Neal admittedly received immunity from prosecution for some crimes, benefits in sentencing for others, or both in exchange for their testimony; and prosecutors arguably forgave Weaver for probation violations in exchange for his testimony.

The Michigan Court of Appeals sidestepped these persistent similarities by focusing solely on the number of witnesses implicating Eddleman. See Eddleman, No. 224957, slip op. at 2. The Fulminante Court, however, rejected such formalistic classification of the type and quantity of evidence, demanding instead that we consider how jurors might view the totality of the evidence remaining when the confession is excised. Fulminante, 499 U.S. at 298, 111 S.Ct. 1246. Even %vith the coerced confession erroneously admitted into evidence, Eddleman’s jury twice informed the court that it was deadlocked before returning its verdict. This datum was not present in Fulminante (in which the Court never mentioned the jury’s deliberations) and provides a powerful indication that, without the confession, jurors probably would view the case as quite weak. We do not believe that a court reasonably could distinguish Eddleman’s case from Fulmi-nante based on the strength of the government’s case.

The other four factors provide equally little basis for a distinction. Regarding the emphasis placed on the confession, the government here, as in Fulminante, made the confession “the centerpiece of this case.” Id. at 297, 111 S.Ct. 1246. The prosecutor in Eddleman’s case referenced the confession throughout his opening and closing statements, called it the “first and foremost” evidence in the case, and told the court that the confession alone would justify a conviction.

Similarly, regarding the relationship between the confession and the other evidence, the prosecutor routinely used the confession to deflect attention away from weak points in his case. He used it to defend the credibility of the four key prosecution witnesses, arguing that their testimony merely reinforced the confession. And he relied on it to undermine the credibility of Eddleman’s core theory that the shooting was revenge for a bar fight unrelated to the activities of the Insane Spanish Cobras, devoting almost all of his cross-examination of the four witnesses supporting that theory to questions about the confession.

Regarding the confession’s evidentiary value, the Fulminante Court distinguished between full confessions and partial confessions. “While some statements by a defendant may concern isolated aspects of the crime or may be incriminating only when linked to other evidence, a full confession in which the defendant discloses the motive for and means of the crime may tempt the jury to rely upon that evidence alone in reaching its decision.” Id. at 296, 111 S.Ct. 1246. Like the defendant in Fulminante, Eddleman gave a full confession, including both a direct admission of guilt and detailed information about the crime, such as what type of gun was used.

Regarding the final factor, the Fulmi-nante Court suggested that the opportunity for cross-examination alone could not overcome the other indications that the admission of the confession was not harmless. See id. at 300, 111 S.Ct. 1246 & n. *588 10. Based on our reading of the trial transcript, Eddleman’s trial attorney also seems to have done a superb job under difficult conditions. In light of Fulmi-nante, however, the Michigan Courts of Appeals could not reasonably have concluded, beyond a reasonable doubt, that his efforts eliminated the prejudice caused by the erroneous admission of the confession.

On each of the dimensions that the Supreme Court identified as relevant to a harmless-error determination, the circumstances of Eddleman’s case parallel the circumstances of Fulminante. No other case-specific factors dictate that the two cases should come out differently. Consequently, we hold that the Michigan Court of Appeals’s harmless-error decision was an unreasonable application of clearly established federal law, as determined by the Supreme Court’s decisions in Chapman and Fulminante.

For the preceding reasons, the district court’s conditional grant of a writ of habe-as corpus is AFFIRMED. The case is remanded to the district court with instructions to order Eddleman’s release unless the state grants Eddleman a new trial within a reasonable period.

. Consider the following exchange during the cross-examination of George Sebastian Pitian:

Q. Would the fact that this man confessed to doing what he thought was a shooting of a gang member, but by mistake hit Joane Georgescu, would that kind of change your mind about that?

A. Yeah.

. We agree with our statement in Nevers that the set of errors that are not harmless under the Brecht standard is a lesser included subset of the set of errors that are not harmless under the standard of Chapman plus AEDPA deference. See Nevers, 169 F.3d at 371-72; accord Polk, 401 F.3d at 264-66 (holding that the state court unreasonably applied Chapman, but that the error was not prejudicial under Brecht); Robertson, 324 F.3d at 307 n. 5 (noting that the alternative of Chapman plus AEDPA deference "is supposed to be more rigorous and less deferential to the state court than the Brecht standard that we re-affirm today”); cf. Gutierrez, 389 F.3d at 305 (observing a "persistent similarity in outcomes” reached under the two standards). Thus, as a practical matter, we think that Brecht and our cases applying the Brecht standard remain helpful even when applying the standard of Chapman plus AEDPA deference.

. AEDPA applies only to claims "adjudicated on the merits in State court proceedings,” and the standard of review it mandates depends on an assessment of an actual decision made by the state court. 28 U.S.C. § 2254(d). Thus, as we frequently have noted and the Supreme Court has confirmed, the Brecht standard continues to apply when federal courts must make a harmless-error determination in the first instance, as when a state court found no error and therefore did not address whether an error would be harmless. See Penry v. Johnson, 532 U.S. 782, 795-96, 121 S.Ct. 1910, 150 L.Ed.2d 9 (2001); Fulcher v. Motley, 444 F.3d 791, 808-09 (6th Cir.2006); Biros v. Bagley, 422 F.3d 379, 388 (6th Cir.2005); Scott v. Gundy, 100 Fed.Appx. 476, 480 (6th Cir.2004); Hill v. Hofbauer, 337 F.3d 706, 718 (6th Cir.2003); Mikesell v. Conley, 51 Fed.Appx. 496, 506 (6th Cir.2002); Stapleton v. Wolfe, 288 F.3d 863, 867-68 (6th *584 Cir.2002). We must scrupulously avoid conflating cases that involve a state harmless-error determination and cases that do not, since AEDPA applies to the former but not to the latter. Cf. Robertson, 324 F.3d at 305 (erroneously stating that, when the Supreme Court relied on Brecht in Penry, a case in which the state court did not consider the issue of harmless error, it "acknowledged the vitality of Brecht's independent harmless-error analysis” in cases when a state court did consider the issue).

. We consider these same factors regardless of the applicable standard of review, which alters only the nature of the bar a petitioner or defendant must clear to demonstrate that an error was not harmless. Compare Stapleton v. Wolfe, 288 F.3d 863, 867-68 (6th Cir.2002) (applying Brecht standard on collateral review) with United States v. Stavroff, 149 F.3d 478, 481-83 (6th Cir.1998) (applying Chapman on direct review); see also Zappulla v. New York, 391 F.3d 462, 468 (2d Cir.2004) (applying the standard of Chapman plus AED-PA deference on collateral review).