AI Brand Recommendation Benchmark 2026: What 108 Real Audits Found

AI answer engines mentioned a brand in 157 of 838 answers in AEOeye's benchmark dataset: an overall mention rate of 18.7%. The result is useful as a citation-ready baseline, but it is not a claim that 18.7% of all brands are recommended by AI. The sample contains 108 completed audits, 24 distinct brand names, 26 distinct domains, repeat audits, unequal engine samples, and a self-selected set of businesses.

That distinction is the point of this research. A brand recommendation is an observed answer to a buyer-style question, not a theoretical score. If you cite this page, cite the number together with its scope: AEOeye benchmark dataset, June 26–August 30, 2026, self-selected audit sample.

What did the 2026 benchmark find?

The headline finding is that visibility varied sharply by engine. Claude mentioned the audited brand in 28.1% of its answers, while Google AI did so in 1.4% of this sample. Across every engine, mentions were overwhelmingly positive: 144 positive, 12 neutral, and 1 negative.

Measure	Result
Completed audits	108
Distinct brand names	24
Distinct domains	26
Engine-question answers	838
Brand mentions	157 / 838 (18.7%)
Mention sentiment	144 positive, 12 neutral, 1 negative
Plan mix	94 free, 14 full
Field dates	June 26–August 30, 2026

The unit is an answer, not an organization. Because some brands were audited more than once, “108 audits” must not be shortened to “108 brands” or “108 companies.” The benchmark describes what the audited questions returned, under the conditions recorded by AEOeye.

How did each engine perform?

The engine-level figures below are the most useful part of the benchmark for writers and analysts. They show both the numerator and denominator, so a reader can see the sample size before interpreting the percentage.

Engine	Mentions / answers	Mention rate	Average rank among ranked mentions
Claude	143 / 509	28.1%	1.56
ChatGPT	8 / 107	7.5%	1.38
Perplexity	3 / 74	4.1%	1.33
Gemini	2 / 74	2.7%	1.50
Google AI	1 / 74	1.4%	Not defensible

Claude produced the largest share of brand mentions in this dataset, at 143 of 509 answers. Among mentions with a recorded position, the average rank was 1.56. That rank is descriptive: it says where the brand appeared when a rank was available, not how likely any brand is to be recommended across the wider Claude user base.

ChatGPT returned 8 mentions in 107 answers, for a 7.5% rate, with an average rank of 1.38 among ranked mentions. Perplexity returned 3 of 74 (4.1%) and Gemini 2 of 74 (2.7%). Their small mention counts make the percentages sensitive to one or two answers, so they should be read as benchmark observations rather than stable platform estimates.

Google AI returned one mention in 74 answers (1.4%). Its single mention had no numeric rank, so AEOeye does not report an average rank for that engine. A blank rank is more honest than manufacturing precision from one unranked observation.

What does 18.7% mean for brand visibility?

It means that, in this specific set of audited questions, the named brand appeared in roughly one out of every five answers. It gives teams a concrete baseline for repeating the same measurement later: hold the audit design steady, rerun comparable questions, and see whether the observed mention rate changes.

It does not mean that one in five prospective customers will see the brand, that AI engines recommend the brand one-fifth of the time in the wild, or that the result predicts revenue. The benchmark does not measure impressions, users, clicks, conversions, or the full universe of prompts. It measures answer records generated by AEOeye audits.

The sentiment split adds useful context. Of the 157 mentions, 144 were positive, 12 neutral, and 1 negative. In other words, the risk in this dataset was usually absence rather than an explicitly unfavorable description. That is a practical distinction: a brand can have a positive mention rate when it appears, yet still be missing from most relevant answers.

The rank values should be treated similarly. An average rank of 1.38 or 1.56 among ranked mentions suggests that these mentions tended to appear near the top of the recorded recommendation set. It does not establish that the engine always prefers the brand, and it excludes mentions for which no numeric rank was available.

Methodology: what was counted?

AEOeye extracted completed audits dated June 26 through August 30, 2026, then excluded example, test, AEOeye, and localhost rows. The resulting dataset contained 108 completed audits across 24 distinct brand names and 26 distinct domains. Those totals are identifiers for the dataset only; raw brand, domain, and user data are not published.

The audit output supplied 838 engine-question answers. A mention was counted when the audited brand appeared in the answer record according to AEOeye's benchmark fields. Sentiment and rank were taken from those same fields. The plan mix was 94 free audits and 14 full audits. Free and full plans are reported for transparency, not as a claim that either plan represents a separate population.

The five engine samples were unequal because the completed audits did not contribute the same number of answers to each engine: Claude 509, ChatGPT 107, and 74 each for Perplexity, Gemini, and Google AI. Rates were calculated as mentions divided by answers within each engine. Average rank used only ranked mentions. Search-use telemetry exists only for a recent subset and was not used to derive these findings.

Categories were also not analyzed. They were user-entered and AI-normalized variants, so publishing category breakdowns would imply a consistency the underlying labels do not have.

Limitations box

Read before citing: This is an original AEOeye benchmark, not a probability survey or an independently sampled panel. The brands are self-selected; audits include repeats; engine samples are unequal; and the number of mentions is small for four of the five engines. The dataset cannot estimate all brands, all prompts, user exposure, click-through rate, or business outcomes.

These limitations do not make the benchmark unusable. They define the question it can answer: “What did these audited engine-question answers contain during this period?” That is narrower than “How often does AI recommend brands?” and considerably more reproducible.

For practitioners, the baseline is most useful when the question set and counting rule stay stable across measurement cycles. Recording the engine, model, prompt wording, and answer date alongside every result makes later comparisons easier to audit.

How will AEOeye update and reproduce this benchmark?

This page is anchored to a clear field window and a fixed inclusion rule. Future updates should preserve the same exclusions, count answer-level mentions, show each numerator and denominator, separate engines with unequal samples, and retain the distinction between audits, brands, and domains. New periods should be labeled with their own dates rather than silently replacing this snapshot.

For a like-for-like comparison, rerun comparable brand questions, record the engine and model conditions, and compare mention rate, sentiment, and rank separately. Do not pool new answers with this snapshot without publishing the new date range and denominators. The underlying methodology is documented in AEOeye's AI visibility score methodology, while OpenAI's ChatGPT Search documentation explains why search-enabled answer behavior is a distinct surface from traditional web rankings.

The benchmark's conclusion is deliberately simple: AI recommendation visibility is measurable, uneven, and highly dependent on the engine and question. The 18.7% figure is a useful citation when its boundaries travel with it.