Methodology: Which brands AI recommends

The questions

Each category has a fixed set of questions. The AI visibility tools category uses 40 questions, all in English and all written the way a buyer looking for a tool would ask (for example, "Any good tools for tracking brand mentions in ChatGPT and Perplexity?"). Every question asks for a recommendation, and none of them contains a brand name. The set did not change in September or October 2026, and we do not change it between editions, because changing the questions would break month-to-month comparison. The full list is in prompts.csv and below.

They are the questions MaxAEO tracks for its own brand in its own product.

The assistants and how answers are collected

We ask each question once a day in each of 8 assistants: ChatGPT, Claude, Copilot, Gemini, Google AI Mode, Google AI Overviews, Grok and Perplexity. That is 320 answers a day when everything works. Answers are collected automatically by MaxAEO's monitoring crawler, using fixed accounts that ask the same questions every day. Most answers are collected between 16:00 and 02:00 UTC (91.4% of collected answers in September and October).

For each answer we keep the answer text, the sources the assistant shows, and a public share link where the platform offers one. Every valid Perplexity, Grok and Google AI Overviews answer has a share link. ChatGPT has one for 94.0% of valid answers in September and 33.0% in October, and Google AI Mode for 8.8% in October (0.1% in September); Claude, Copilot and Gemini have none. For Google AI Overviews and some Google AI Mode rows, the link is a Google search URL, so it shows that search as it is today rather than the answer we recorded. Where there is no share link, write to nancy@maxaeo.ai for the answer text to check a row.

Which answers count

Because of that last row, the status column in the downloads adds up to 520 failed and 14 no_ai_overview in September, and 1,730 failed and 6 no_ai_overview in October. Valid answers are not affected. The reason is under Known biases.

An answer is invalid if it is shorter than 120 characters, or if it is an error message instead of an answer (for example Gemini's error screen, ChatGPT's "Something went wrong while generating the response", "couldn't load", or a broken map tile). For Google AI Overviews, a search where no AI Overview was captured is recorded as "no AI Overview" and is not a valid answer. All of these rows are kept in the downloads with their status, so nothing is silently dropped. Only valid answers are used in the rankings.

Gaps by day

The full list, by day, platform and question, is in missing.csv. The largest gaps:

• 2026-10-06: 17 of 320 answers were valid, the worst day in either month.
• 2026-10-01 to 2026-10-05: 154 to 164 of 320 valid each day.
• 2026-09-30: 172 of 320 valid.
• 2026-10-09: 192 of 320 valid; 2026-10-10: 208 of 320 valid.
• 2026-09-24: Copilot, Google AI Mode, Google AI Overviews and Grok were not queried at all that day.

Most of these gaps hit the same assistants at once (Gemini, Google AI Mode, Google AI Overviews and Copilot), which points to a problem on our collection side on those days rather than at the assistants. See "Known biases" below.

Every day, as recorded in the downloads. Shaded days are the days with the large-scale collection failure described under Known biases.

When a brand counts as "named"

We only count brand names written in the answer text. We match names with a public alias list of 44 brands (brand-aliases.csv). Product lines are merged into their company: for example, "Semrush AI Visibility Toolkit" counts as Semrush. Brand names that are also ordinary English words (Profound, Conductor, Frase, Goodie, AthenaHQ) are matched case-sensitively to avoid false hits.

Web addresses inside the answer do not count as naming a brand. Some assistants, Google AI Mode in particular, embed source links in the answer text; we remove link targets and bare URLs before matching, so a link to maxaeo.ai is not counted as naming MaxAEO. Source labels do not count either: Google AI Overviews, Google AI Mode and some ChatGPT answers show where a paragraph came from as a short line such as "HubSpot +1" or "maxaeo.ai +2", and we remove those lines before matching.

We do not use any model-based judgment of whether a brand was "mentioned". Everything on the ranking can be recomputed from the public CSV.

Swipe the table sideways to see every column.

Rankings and other measures

• Named rate: valid answers that name the brand, divided by all valid answers.
• Named first: share of valid answers in which the brand is the first of the 44 brands to appear.
• Average position: among answers that name the brand, the average order in which it first appears (1 = named first).
• Days in the top 3: for each day, brands are ranked by how many valid answers named them; we count the days a brand was in the top 3.
• Change vs last month: the named rate this month minus the named rate last month, in percentage points.
• Own-site citation rate: among answers that name the brand and cite at least one source, the share that cite the brand's own website.

Overlap between lists

To measure how much two brand lists agree, we use the size of the intersection divided by the size of the union, averaged over all pairs. Pairs where either list is empty are skipped.

• Between assistants: the same question on the same day, compared across every pair of assistants.
• Day to day: the same assistant and question on two consecutive days.

We report each overlap two ways. The first uses only the 44 brands on the alias list, and you can recompute it from the CSV: 51.4% day to day and 34.2% between assistants in October. The second uses every brand name our monitoring system extracted from the answers, normalized for spelling and product lines (1,189 brands after normalization): 46.8% and 28.9%. The second version counts long-tail names that change often, so it comes out about 5 points lower. We treat it as a lower bound, because spellings that our normalization did not merge also push the overlap down.

Known biases

• We are a participant. MaxAEO publishes this report and is one of the 44 brands. We also publish articles on the topics these questions ask about. We do not change MaxAEO's position or highlight its row; the data is downloadable so you can check.
• Fixed accounts. Answers come from accounts that ask the same questions every day. Some assistants may personalize answers for these accounts, so what we see may differ from what a new user sees.
• Collection failures on some days. On 11 days in October (2026-10-01 to 10-06, 10-09, 10-10 and 10-28 to 10-30) and on 2026-09-30, collection for Google-related assistants and Copilot failed on a large scale. On those days, valid answers were 9.0% for Gemini (93.0% on other days), 32.5% for Google AI Mode (97.9%) and 26.7% for Copilot (84.5%). On the same days Google AI Overviews mostly came back empty, and our rules record an empty result as "no AI Overview". We think most of these are collection failures rather than Google not showing an Overview: 345 of October's 351 "no AI Overview" rows fall on those days, and on all other days an AI Overview appeared in 98.8% of searches in September and 99.2% in October. We therefore record those rows as failed collection and do not report the AI Overview appearance rate as a trend. Our collection team traced this to a network fault followed by a change of network route: Google often served those requests from its Hong Kong site, which returns no AI Overview. The Copilot gaps were mostly accounts running out of quota.
• Uneven platform mix. Because the gaps are concentrated in a few assistants, October has far fewer valid answers from some assistants than others (668 from Copilot, 1,190 from Claude). Overall named rates are pooled across assistants, so an assistant with more valid answers weighs more. The by-assistant table on the category page shows each assistant on its own.
• A change in our collector in September. From 2026-09-17, answers on every assistant took noticeably less time to collect (for Claude, the median fell from 55.3 to 21.4 seconds), which suggests something changed in our collection setup that day. We did not find a matching jump in brand results on or right after that date.
• Limited brand list. Rankings cover the 44 brands on the alias list. Brands outside the list are not ranked, and overlap measured on the 44 brands is higher than overlap measured on every name (see above).

Data sources and credits

• Answers: collected by MaxAEO's monitoring, as described above.
• Day-to-day overlap comparison: Source: Namebeam Edition 1 (DOI: 10.5281/zenodo.23626831), licensed CC BY 4.0. We quote its 12%–40% consecutive-day overlap range for context only. Its category, collection method and brand set differ from ours.

Downloads and license

Each edition has its own set of files. Files from earlier editions stay at the same address. All files are listed on the AI visibility tools page.

• answers-YYYY-MM.csv.gz: one row per attempted answer. Columns: date, platform, question ID, answer ID, status, brands named in order, number of brands named, whether MaxAEO is named and at what position, cited domains, cited URLs, public share link.
• Answer text: not a download in this edition, because of its size. Each answer in the CSV has a public share link where the platform offers one, and we send the raw text on request.
• prompts.csv, brand-aliases.csv, missing.csv, summary.json.

The compiled data and every column we derived (statuses, brands named, positions, cited domains, summaries) are licensed under CC BY 4.0. Please credit "MaxAEO, Which brands AI recommends" with a link to this page. The answer text is output from each AI provider. We share it only so you can check our counts; we do not license it.

Corrections

If you find an error, write to nancy@maxaeo.ai. We add a dated correction note to the edition it affects and do not overwrite earlier numbers.