Which Odyssey translation wins a blind reading test?

Finding: Fagles led this test. Robert Fagles’s Penguin translation ranked first among the eight tested translations in the 674-ballot published analytical cohort. It led the raw results, the regularized pairwise model, the scene-adjusted model, and all 50,000 whole-ballot bootstrap resamples. This is strong evidence about reader preference within this study but not a universal verdict on the best Odyssey translation.

This experiment measures relative reader preference among eight translations, across a study panel that ultimately included eight short scenes, in a self-selected and mostly American social-media sample. It does not measure fidelity to the Ancient Greek, the experience of reading an entire translation, or the preferences of all Odyssey readers.

Answer: the eight translations with the highest estimated named search attention among 17 candidates.

We evaluated 17 candidates using trailing 12-month U.S. search-volume estimates for exact full-name-plus-Odyssey queries and capped the test at eight to keep the comparison task manageable. This is a reproducible measure of named search attention—not sales, readership, classroom assignment, or literary importance. The report ranks these eight translations, not every English Odyssey.

Sample: a mostly U.S., mobile, social-media convenience sample recruited during attention around Nolan’s Odyssey.

The largest traffic source was our Instagram post. TikTok, Facebook, YouTube, and readers sharing the test also contributed.

The convenience sample came from the cultural moment that inspired the test: Christopher Nolan’s Odyssey. Social media reached people already interested in the film and Homer. That audience interest is the study’s primary selection bias.

Device, geography, and source figures come from campaign and visitor analytics. Gender comes from paid campaign-delivery data and was not collected on the ballot. These figures describe the audience reached by the campaign; they cannot be mapped precisely to either the 685 counted completions or the 674-ballot published analytical cohort. We did not retain raw IP addresses or direct identifiers; one-way IP hashes were used to detect retakes.

Finding: named search attention was a weak proxy for preference in this eight-translation comparison. Emily Wilson had the highest named search attention but did not lead the blind test.

Emily Wilson argues that readers may mistake archaic language for greater authority or authenticity. Here is her full point:

“Mild stylistic archaism is often accepted without question in translations of ancient texts and can be presented as if it were a mark of authenticity. But of course, the English of the nineteenth or early twentieth century is no closer to Homeric Greek than the language of today. The use of a noncolloquial or archaizing linguistic register can blind readers to the real, inevitable, and vast gap between the Greek original and any modern translation.”

Below, we compare publication date with blind reader preference. Newer translations look stronger across the full 300-year span. Remove Pope and the pattern weakens. Among translations since 1960, it mostly disappears. This does not refute Wilson’s point that readers may be biased toward “mild stylistic archaism,” but it suggests they were not reflexively choosing the most archaic English in this test.

Across the eight translations, Google AI mentions had a weak descriptive relationship with blind preference (correlation 0.22). Wilson appears most often in DataForSEO’s observed Google AI corpus but ranks sixth in pairwise strength.

This relationship is also weak (correlation 0.32). ChatGPT surfaces some leaders, but Mendelsohn ranks third in the blind test and appears rarely.

Answer: very few, and it did not matter. We excluded 11 timed completions under 90 seconds. Fagles still leads when they are restored.

Finding: Fagles still leads. Restoring all 11 completions under 90 seconds leaves the pairwise ranking unchanged. Raw match shares also move very little; Fitzgerald and Wilson tie at 79 top matches.

This is just for auditing purposes. The original ranking is more robust because it applies the report’s consistent 90-second quality threshold. The zero-discard view shows that the winner was not created by that cutoff.

We found no coordinated bot or voting campaign. The site counted one ballot per anonymous browser and blocked retakes and later submissions from browsers that had already cast a counted ballot.

To check for coordinated activity on shared networks, we converted each IP address into a private one-way hash: a nonreversible fingerprint that lets matching addresses be compared without keeping the raw address in ballot records. Retakes never entered the public totals.

To select the eight-scene panel, we used GPT-5.6 Sol to construct a fixed pool of 25 candidate scenes using only Samuel Butler’s public-domain Project Gutenberg translation. Candidate scenes were ranked by their recurrence in nine study guides under the assumption that well-known plot points would increase test completion rate. GPT-5.6 Sol then narrowed that fixed pool to eight scenes selected for narrative and semantic breadth across the poem.

The panel-selection input did not contain any of the seven other tested translations or any reader results. It therefore could not directly optimize for wording that favored any competing translator. We evaluated the selected panel against all 1,081,575 possible eight-scene subsets of the same 25 candidates. Just 5,059 subsets, or 0.47%, matched or exceeded it on both recognition and semantic breadth. This is a benchmark to defend against cherry-picking. It doesn’t mean scenes were randomly sampled.

Fagles leads overall, but other translators win individual scenes. Pick one to compare its first- and eighth-place passages.

Finding: not much. We searched for two to five recurring reader groups—such as one camp favoring literal translations and another favoring more literary ones—but found no division that was both stable and meaningfully distinct.

The two-group solution was highly repeatable, reappearing in 97.6% of resamples, but its measured separation was 16.4%, below our stated 20% qualification threshold. We therefore treat it as evidence of modest recurring taste variation rather than two sharply distinct camps.

That does not mean readers agreed on every passage. It means the overall result does not appear to be an artificial average produced by several large, opposing taste camps.

We are not classicists nor do we know Ancient Greek, so we’ve provided some essays below. From what we’ve learned, the question isn’t so simple. Every translation necessarily interprets and transforms the original. Each choice preserves some features while changing, narrowing, or sacrificing others. Here is a simple three-part model that begins to explain some of the difficulty.

“My goal is not to rank them, nor to criticize my esteemed predecessors and colleagues.” - Emily Wilson

Wilson’s essay instead compares how particular translations echo and depart from the Greek. This is not our expertise. If it is yours, follow the sources above and judge the arguments for yourself.

As of July 22, 2026, Scrivium has no financial or editorial relationship with any tested translator or publisher. We do have a commercial interest in attracting readers to our interactive mythology course, but we do not think that shaped the result.

We hid translator names, dates, and publishers. Readers were shown up to 12 comparisons. For 668 of the 674 published ballots, the opening eight formed a balanced cycle: every available translator appeared exactly twice. Before each of the final four comparisons, the test refit a regularized Bradley–Terry model to that reader’s previous non-skipped choices. It first prioritized translators with the least evidence and, when skips had fragmented the comparison network, pairs that connected those fragments. Among the eligible pairs, it favored strong contenders with similar scores, greater uncertainty, and no previous head-to-head meeting. Scene assignment followed the phase-specific rules detailed below. Left and right positions were independently randomized for every comparison. Skips changed neither translator’s preference score, although the pair and scene still counted as having been shown. Detailed implementation facts:

• Initialization: Every translator began at ability 0, equivalent to a neutral 50% predicted win probability.

• Updating scores: After each non-skipped choice, the browser refit the model from all previous choices. It ran 80 gradient iterations with L2 regularization of 0.7, shrinking sparse estimates toward zero. Skips were omitted from the ability fit.

• Eligible pairs: The starting universe was every unordered pair among the translators present in that ballot—28 pairs for eight translators.

• Coverage and connectivity filters: If non-skipped choices left disconnected groups, only pairs joining two groups were considered. It then retained pairs containing the largest possible number of translators tied for the fewest decisions.

• Repeated pairs: Previously shown pairs were excluded whenever an unseen pair survived those filters. Repeats were permitted only when necessary and were penalized according to 1 / (1 + previous meetings). A skipped pair still counted as previously shown.

• Closeness: For abilities aᵢ and aⱼ, closeness was 1 / (1 + |aᵢ − aⱼ|). This was only one part of the composite score, alongside contender strength, uncertainty, connectivity, and novelty.

• Final selection: The algorithm kept the six highest-scoring eligible pairs and made a weighted selection among them. Higher scores were more likely, but the highest score was not automatically chosen.

• Scene assignment: The method changed prospectively. Published ballots 1–6 used the earlier four-scene cycle. Ballots 7–447 balanced scene exposure while also considering editorial strength, lexical contrast, and pair freshness. For published ballots 448–674, the selector removed the immediately previous scene when alternatives existed, retained the least-shown scenes, sorted them by ID, and selected one with a ballot-specific deterministic hash. Only this final phase was fully blind to candidate identity, wording, and the reader’s selections.

• Unavailable material: Pair selection only saw translators in the session’s candidate set; scene assignment only saw passages present in that session. Nothing unavailable was imputed or backfilled.

• Left/right order: The logical pair was independently flipped with a 50/50 cryptographic random draw before every comparison.

• Randomness and seeds: There was no fixed study-wide assignment seed. The server cryptographically shuffled translator and passage order and generated a random ballot UUID. That UUID became a per-ballot seed for reproducible adaptive-pair and scene assignment. Left/right flips remained separately cryptographically random and unseeded.

Also note, a JSON validation issue was corrected prospectively. Daniel Mendelsohn was unavailable for the first 6 counted ballots; all six were retained, and he entered on ballot 7. Harmonious household and The Lotus-Eaters were unavailable for the first 457 counted ballots, of which 447 remain in the published cohort; they entered on counted ballot 458, or published ballot 448. No ballots were rewritten or excluded because of these corrections. The report separately excludes 11 ballots for finishing under 90 seconds; 10 of those ballots were collected during the scene-affected period. Opening rounds balanced the options then available, while the scene-adjusted model controls for translator-by-scene and recruitment-wave effects.

The published ranking fits a regularized Bradley–Terry model:

The compensated scene model adds partially pooled scene effects, recruitment wave effects, and a first-position term:

The 95% ranges come from 50,000 whole-ballot bootstrap resamples, preserving the dependence among one reader’s choices. Published charts use 674 ballots and 7,919 non-skipped choices. Retakes and later writes from the same anonymous browser never entered the public totals.

The frozen snapshot, analysis settings, downloads, and reconciliation checks are below.

For interviews, data and methodology questions, publication-ready charts, or factual corrections, contact Kyle Cureau at kyle@scrivium.com or text Kyle at 415-390-6475 and include your outlet and deadline.