How the AI Answer Corpus is built

This page is the instrument. It exists so that any figure on the study pages can be checked against how it was collected, and so that someone with their own prompts can rebuild the same measurement.

The instrument

Each prompt is sent, unchanged, to two engines through DataForSEO's LLM Responses endpoint with web search on: ChatGPT as the API model gpt-5-mini, and Perplexity as sonar. Both models are pinned, so a change in an answer is not a change of model. Each prompt is sent 12 times per monthly fill. Every response is stored whole, with its source list, and parsed by the same code that produces the tracker's client reports, so a study cannot disagree with a report about whether a brand was named or cited.

Two fills are in the corpus: 12 to 26 September 2026 and 1 to 5 October 2026. The calls carry no location, account or chat history. The consumer ChatGPT and Perplexity apps add all three, so these rates describe the engines' retrieval and naming behaviour, not what any one user sees.

The sample

A cell is one prompt, in one month, on one engine: 12 runs. The 12 tracked sites span 7 industries: legal (two practice areas), energy B2B, marketing and SEO, a consumer supplement, a technology service, and advisory. Two sites ran in one month only, so 186 prompts have answers in both fills; the cross-month comparisons use those.

The prompts

Prompts are phrased the way a buyer would ask an assistant, not as keywords, and were written per tracked site, so the corpus is not a sample of search demand. 230 of the 331 ask for a recommendation (which firms, best providers, who should we hire); the rest ask for information. One 66-prompt energy panel was written for no tracked brand and ran in September only. Prompts that name a brand are flagged and the study pages say when they are excluded.

Sources and the three bases

ChatGPT returns a source list of 4.3 to 4.5 URLs on average when it searches, and 32.7% of its runs return no sources at all. Perplexity returned 19.9 sources per answer in September and 10.0 in October, when the API moved from 20 to 10; in October exactly ten came back in 98.0% of answers. Every citation figure therefore names one of three bases:

• ChatGPT, every URL it returned.
• Perplexity, first 10: the basis the two fills share, and the default on every study page. It assumes the list order is a ranking, which is only partly true: a third of September's in-text citation markers point past position 10.
• Perplexity, all sources: September's full lists of 20, used only where a page says so.

Two ways of counting appear on the study pages. Composition is the share of answer-host pairs (one host cited in one answer is one pair). Reach is the share of an engine's answers citing a type at least once. Reach is read within an engine, never across, because ChatGPT's no-source runs lower every one of its reach figures.

Timing

The 12 repeats of a cell are not all sent at once. The span from the first to the last run of a cell has a median of 118 minutes, a 90th percentile of 23 hours and a maximum of 46 hours. Study 4 tests whether run order predicts the answer; it does on Perplexity and not on ChatGPT.

Statistics

Every rate on the study pages carries a 95% interval where the sample allows one. Intervals on shares of answers or citations are bootstraps clustered by prompt (400 to 4,000 draws depending on the page), because 12 runs of one prompt are not 12 prompts. Single proportions use Wilson intervals, differences of proportions use Newcombe intervals, and the run-order and stale-search tests use within-cell permutation (2,000 shuffles). Set overlap is Jaccard; the difference between two source mixes is total variation distance. Where a figure is a median across prompt-months, its bracket is an interquartile range and the page says so.

Each finding carries one of two grades. Strong: the figure held under every check its study ran (a different basis, a month at a time, a hand-checked sample, a placebo). Directional: the direction held but the size rests on a rule, a keyword pattern or a small slice, and the page names which. Findings that failed review are not on the pages; a reviewer re-derived every published figure from counts before publication, and nothing on these pages is estimated or interpolated.

What is withheld, and why

• Client names, site names and per-site figures. The sites pay to be tracked and their prompts are commercial; every published figure is pooled across the 12 sites.
• Per-industry figures. Three of the 7 industries hold a single tracked site, so an industry figure would be a client figure.
• The dataset itself. There is no download or export. The page is the citation, and the tables on each page are described well enough to rebuild the instrument on your own prompts.
• Findings confounded by crawler access. Part of the sample sat behind a web application firewall that blocked AI crawlers during both fills. Findings that depend on whether an engine could read a site were held back until that is resolved.

Refresh

The corpus grows by one fill a month. The study pages refresh after the November 2026 fill closes, about 15 November, when a 15-prompt research panel joins the tracked prompts: questions chosen for what these pages raised, not for any client. Figures on a refreshed page carry the new as-of date, and a figure that moves by more than its interval is called out on the page rather than overwritten silently. The house rule for every number we publish is on belief against measurement; the same discipline applied to thirty years of SEO data is on what the ranking factor studies measured.

Want a question on the November panel? Send the prompt, phrased the way a buyer would ask it.

Quoting a figure? Link to this page as the source: https://www.bestseopodcast.com/research/ai-answer-corpus-method. Figures are dated 2026-10-25 and cover answers collected 12 to 26 September 2026 and 1 to 5 October 2026; the next refresh is after the November 2026 fill closes, about 15 November. The page is the citation: there is no download.

Who did this

Matthew Bertram

Author · Host of The Best SEO Podcast since 2009 · Owner and CEO, EWR Digital

Matthew Bertram runs EWR Digital, the SEO and AI-visibility agency whose monthly tracker produced this corpus, and he is the author of LLM Visibility: A Decision-Grade System for Winning AI-Mediated Discovery. The studies were built, re-derived from counts and reviewed before anything was written up; every figure on this page traces to a script in the project record. Questions about the method go to him through the contact page. More about Matthew Bertram.