Method and raw data

• Each of the ten questions was typed as written into each app, in a fresh chat or search, and the answer was read from the page and saved with a screenshot and its links.
• ChatGPT: chatgpt.com home page, logged out, fresh chat per question. The default experience is used with no model or effort picker (none is shown to logged-out users); the composer sends the question as-is, and ChatGPT decides whether to search the web. off (no account)
• Gemini: gemini.google.com/app, logged out, fresh chat per question off (no account)
• Google AI Overview: google.com/search (All tab) for the AI Overview; udm=50 for AI Mode; signed out
• Google AI Mode: google.com/search (All tab) for the AI Overview; udm=50 for AI Mode; signed out
• Location: United States, residential IP, America/New_York time zone, en-US. Google requests carry hl=en and gl=us.
• One run per question per surface (run 1). AI answers vary between runs; one run is an observation, not a rate.
• Apps can personalise answers and vary them from one ask to the next. Each cell is one observation.

• Mentioned, Recommended and Linked to your site come from fixed rules (name matching, cue words and web-address matching, listed in targets.yaml). No language model judged any cell.
• "Recommended" is a cue-word check (recommend, best, top pick or similar next to the name, or a list under a recommendation heading). It can miss or over-flag a recommendation, and it never infers a rank from list order.
• "Linked to your site" means the answer carries a link to the brand's own web address.

Checked 2026-10-31T15:39:59Z. one GET each for robots.txt, the sitemap and the homepage, with a neutral browser-style user agent; no crawler user agent is spoofed. robots.txt returned HTTP 200. Saved robots.txt, crawler.json.

• Sitemap: https://plausible.io/sitemap.xml: HTTP 200, 167 URLs, declared in robots.txt
• Home page: HTTP 200, canonical https://plausible.io/, meta robots None, X-Robots-Tag None

This reads the rules a site publishes. It does not prove what a crawler fetched (that needs server logs), and a CDN or firewall can still block a crawler robots.txt allows.

Before the app capture, the same ten questions were also sent to the OpenAI and Gemini developer APIs with web search on, twice each. This is a different product from the apps: the apps add their own instructions, memory, personalisation and interface. The numbers on the main page never include it. In that test Plausible was mentioned in 36 of 39 answers and its site linked in 29; 1 of 40 calls failed (shown as failed).

• OpenAI API, run 1 (gpt-6-luna, 2026-10-31 15:41 UTC): Q1, Q2, Q3, Q4, Q5, Q6, Q7, Q8, Q9, Q10
• OpenAI API, run 2 (gpt-6-luna, 2026-10-31 15:43 UTC): Q1, Q2, Q3, Q4, Q5, Q6, Q7, Q8, Q9, Q10
• Gemini API, run 1 (gemini-2.5-flash, 2026-10-31 15:41 UTC): Q1, Q2, Q3, Q4, Q5, Q6, Q7, Q8, Q9, Q10
• Gemini API, run 2 (gemini-2.5-flash, 2026-10-31 15:43 UTC): Q1, Q2, Q3, Q4, Q5, Q6, Q7, Q8, Q9 failed (HTTP 429: { "error": { "), Q10

Estimates from each response's usage fields at list prices (OpenAI and Gemini pricing pages, read 2026-10-31). Gemini grounding is counted at the list price of $35 per 1,000 grounded prompts even though the API documents a free daily allowance, so the figure is a worst case. Not an invoice. A one-question setup check on both engines ran first ($0.059) and is not part of the sample; total spent on this page's API data, setup check included: $1.259. Every API answer links to its stored raw response; each raw file is checked against its SHA-256 when the page is built: 40 of 40 verified.

• Ten questions, one category, one ask per question per app: enough to show the method, too few to rank brands.
• A change after a website edit counts as real only if it is larger than the normal movement between identical asks. A proper baseline is five or more asks per question per app with frozen questions, and dated change notes next to them.
• No traffic data. Referral traffic from AI products would come from analytics and server logs, with traffic of unknown source kept in its own bucket.

consumer.jsonresults.csvresults.jsonrun-log.jsonlquestions.yamltargets.yamlrobots.txt (saved)crawler.json

consumer.json holds every app answer; results.csv and results.json hold the separate developer-API test; the run log keeps every attempt and error.