Blog · 24 September 2026 · By HeardOf

Which AI-visibility studies could you check? Sample, prompts, raw data

Which of the 61 study-shaped vendor posts are studies at all?

Thirty-three of them. The 61 are every post on the Profound, Otterly, Peec and AthenaHQ blogs whose slug carries study, experiment, report, analysis, benchmark, data or research as a whole word; the rule and the counts — Otterly 36, Profound 15, Peec 7, AthenaHQ 3, Geoptie 0 — come from our census of the five blogs, and on 24 September 2026 we opened all 61, rendered each in a browser, and read them. Thirty-three report a measurement of their own: 17 studies of data the vendor holds, 12 experiments, two guest experiments measured with the vendor's tool, and AthenaHQ's two summaries of its own report. The other 28 are seven case studies, six guides, nine product announcements, two company announcements, a methodology page, a category index page the sitemap lists as a post, a third party's study retold, and a recap of five other studies.

The census warned that a slug is not a reading, and this is what the reading finds. Profound's 15 study-shaped slugs hold three studies; its /blog/research address is the index of its Research category, and the word data sits in the slugs of a data-nodes launch, a prompt-recommendation launch and a Reddit study alike. Otterly's 36 hold 24 measurements, Peec's seven hold four, AthenaHQ's three hold two.

How many AI-visibility vendor studies state their sample size?

Thirty-one of the 61 print a figure for what they measured, and 28 of the 33 measurement posts do. The three others with a figure are two case studies and a retold study: NOLA Marketing's 64 prompts, Stella Rising's "4,500 prompts and 1.8 million citations" — an agency's analysis on Otterly's data, told in an interview of 31 July 2026 — and the 150 questions of the komm.passion study of 21 January 2026 that Otterly retells. The figures run from seven prompts, in Otterly's hidden-text experiment of 9 April 2026, to "over 4 billion AI citations and 300 million answer engine responses" in Profound's Reddit study of 10 November 2025, with 20,092,769 URLs in Otterly's dead-citations study and 232,000 citations across 13,000 listicles in Peec's listicle analysis of 3 March 2026 between them.

Five measurement posts print no figure. AthenaHQ's two report summaries say "millions of AI-generated responses across B2B and B2C"; the only count of what was measured in either is "over 500 conversations" — demos, strategy sessions and onboarding calls — for a supplementary analysis. Otterly's ChatGPT ads study counts 16 industries and the 50 brands advertising most, and never says how many prompts or answers it read. Its Markdown-versus-HTML experiment of 1 April 2026 describes two page pairs and no run count, and its taste experiment of 11 June 2026 fielded five image pairs to "a panel of hundreds of respondents". A figure describing the product rather than the study — Profound's "dataset of 100M+ answer engine prompts", in a feature launch of 15 July 2025 — is not counted for anyone.

Do the studies say when the data was collected?

Twenty-one of the 61 name at least a month, and 20 of the 33 measurement posts do. The other 13 give a duration with no calendar on it, or no window at all. Two Otterly studies, of 7 and 14 May 2026, analyse the same 1,028,959 URLs "over a 24-hour window", and neither says which day; its dead-citations study covers "one full month", its YouTube study of 2 March 2026 "a 30-day period", its llms.txt experiment of 5 February 2026 "a continuous 90-day period", its Markdown experiment 14 days, its image-metadata experiment of 21 April 2026 one week and its taste experiment "several days", none anchored to a date. Its State of AI Search study of 29 April 2025 and its hidden-text experiment give no window, and its page-deletion experiment shows its dates "as day numbers relative to each removal". Peec's 30-million-source domain analysis of 31 March 2026 carries no period at all, and neither does Profound's intent study of 25 June 2025, of "50M+ ChatGPT prompts". Three more give a month and no year — Otterly's Claude study, "in June", with "a short window from June 9 to June 13"; its glossary experiment, "21 August to 8 September"; its year-in-title experiment, "June 25 to July 9" — and we counted those as dated, since each post's own date settles the year, and flagged them.

Where a window is printed, it is a day or month range: "12 May to 26 August 2026" for the livestream experiment, "from July 2 to 15, 2026" for the ads study, "between January 1 and June 1, 2026" for the LinkedIn study of 3 June 2026, "from September 2025 through January 2026" for Profound's shopping analysis of 3 March 2026, "between October 2025 and January 2026" for Peec's fan-out analysis of 12 February 2026. Seventeen of the 33 measurement posts state both a figure and a month.

Does any vendor publish the prompts it ran?

None of the 33 measurement posts prints its full prompt set or links a page that does, on 24 September 2026. Four print an example: Otterly's hidden-text experiment gives one of its seven ("What text do you see on https://thebluehaiku.com/?"), its awards experiment of 25 May 2026 one example per category for six of its 32, its gender study of 5 June 2026 the three versions of one of its 42 questions, and the TRYSEO video experiment of 22 April 2026 "Examples of the prompts in scope". Outside the measurements, a tutorial of 3 May 2024 prints the one ChatGPT prompt it teaches, Peec's brand-mention-gap guide of 5 February 2026 prints the prompt in its worked example, and Otterly's Bacula case study of 16 June 2025 names two queries it tested.

The other 29 print none; where they mention the set, they describe it — "a fixed set of non-branded, software review prompts" in Peec's listicle analysis, "a consistent set of commercial prompts for each of the 16 industries" in Otterly's ads study, "a broad, continuously monitored set of SaaS and tech prompts" in its Claude study, "125 tracked prompts with language: German, region: Germany" in its livestream experiment, "30 healthcare-related queries per product, totaling 150 questions" in the komm.passion study it retells. A prompt a post prints as an illustration of a type — Profound's "best business laptops with strong battery life", Peec's laptop fan-out example — is not counted as published, because the post does not say it was run.

Can you download the data behind any of them?

No. None of the 61 links a download, a public spreadsheet or a repository holding the per-unit data — per prompt, per answer, per citation, per URL — behind its figures, as served and as rendered on 24 September 2026. What exists instead is worth listing, because each is the thing a reader would click expecting data.

Otterly's State of AI Search study of 29 April 2025 and its AI Citation Economy report of 1 February 2026 each carry a "Download the full … here:" line followed by an embedded form; no file address is in either page. The 2026 report also links "Explore Interactive Dashboard", a page that shows citation shares by category and, on a toggle, the 500 most-cited domains with a count for each, filterable by engine and country, and says "Data is from September 2025". Totals per domain, not rows per citation or per answer. Otterly's methodology page of 1 June 2026 links a public Google Sheet, "Public GEO Experiments - OtterlyAI", which on 24 September 2026 held 127 numbered rows, 117 of them naming an experiment, with columns for start and end month, tactic, type, difficulty, status, hypothesis, a one-line result, the researcher and the post's address. It is a list of experiments; it is not the data behind any of them. Three experiment posts link the same sheet through a "GEO experimentation sheet" page. AthenaHQ's two summaries link "the full report", a form page whose HTML names a PDF that answered directly to a request on 24 September 2026 — a 29-megabyte report, not a dataset. Profound's Gartner post links a report for Gartner clients, and its Prompt Volumes post of 25 July 2025 says results "can be exported as a CSV" inside the product.

What does HeardOf publish, by the same four questions?

Two of our own posts on the run of 4 September 2026, read live on 24 September 2026 and asked the same four questions, score yes, yes, yes and no. Who the AI actually cites states "20 prompts, 40 answers, 564 citations, 463 distinct pages, 269 distinct domains" and the day, 4 September 2026; named but not cited states "240 brand-answer pairs — 24 brands, 10 answers each" from the same run, re-counted on 23 September 2026. Who the AI actually cites says its 20 prompts are the four non-dental blocks of the 25 printed verbatim in our post of 4 September 2026, under the line "Here are all 25, verbatim."; named but not cited counts the same run and links both posts. That is the full set, by link, which is the cell no vendor measurement fills.

The fourth cell is a no. Neither post links a download, a sheet or a repository of the stored answers; the external links on the first are two arXiv papers, a Google announcement, an OpenAI announcement and the Trellner study, and the second has none. Both print their counting rules in prose. The test this post applies is the line Rand Fishkin wrote in SparkToro's post of 28 January 2026 on the study he ran with Gumshoe's Patrick O'Donnell — "stats-backed, publicly-reviewable research", read on 24 September 2026 — and by it our two posts pass on the prompts and fail on the data. We are printing that beside the 61 rather than above them. The same raw-data question, put to the studies of how often AI assistants get facts wrong, is answered in our compilation of accuracy studies.

How was this measured?

The set is the 61 study-shaped slugs of our census of 23 September 2026, re-derived from the same 574 records with the census rule, which returned the same 36, 15, 7, 3 and 0. On 24 September 2026 each post was fetched with curl and a browser user-agent, then rendered in Chrome 153 headless: a nine-second wait, a scroll, the page's text as it opens; then every collapsed details element opened and every tab, accordion and expander outside the navigation, header and footer clicked, and the text read again. No post drew anything on click that changed a cell. Every quoted span was found in both the served text and the rendered text, and the totals above were re-derived from the per-post cells by script rather than added by hand.

The four rules, so this AI search study methodology can be re-run on the same pages. A sample size is a figure with digits for the units the post's own measurement covers — prompts, answers, citations, URLs, pages, videos, accounts; "millions" is not one, and neither is a dataset size that describes the product, nor a count of industries alone. A date is at least a named month; a duration with no calendar anchor is not, a bare year is not, and the post's own publication date is not. Prompts are published when the full set is on the page or on a page it links, printed as an example when at least one from the set is, and not published when the set is described or the printed prompt is a template or an illustration. Raw data is a download, a public sheet or a repository with per-unit data; a report PDF, a form, an interactive page of shares and per-domain totals, a list of experiments and a product export are recorded in the notes and not counted. Forty-four of the 61 were published more than 90 days before the read and are labelled historical; the read is of the page as it stood on 24 September 2026, not of the study when it ran. We did not ask any vendor for data that is not on the page: a vendor that shares data on request scores the same here as one that does not, because the test is of what is public.

Everything this post says about the 61 posts is a reading of the pages in the tables above and the four linked pages named in the download section, made on 24 September 2026, and you can open any of them from the vendors' blogs; everything it says about our own two posts links them, and they were read live the same day. What the closing says about the audit we sell comes from our own post on which engines we check, as read on 24 September 2026, and from nothing newer.

How often do AI assistants get facts wrong? Seven accuracy studies, compiled with their units