42132
ChatGPT visitors a day, 22–28 Sep to 26–29 Oct. 3.1× in five weeks.
- 54%
- of all visitors came from chatgpt.com, 22 Sep to 31 Oct
- 2 of 8
- ChatGPT answers recommended it in our 7 Nov check, and 3 of 8 linked it
What AI says about you. With the receipts.
Forthsay asks the models behind ChatGPT, Claude, Gemini and Perplexity how they describe and recommend your site, shows the raw answer behind every number, and helps you publish facts they can check.
Asked through each vendor’s official API, every week
- ChatGPT
- Claude
- Gemini
- Perplexity
gpt-6-luna · claude-haiku-5-5 · gemini-3.5-flash-lite · sonar
Sample data
34%picked as a choice, week of 2 Nov41 of 120 answers · 95% range 26% to 43%
+8.3 pts after the change, range 2.7 to 13.9: up, the range agrees
- ChatGPT
gpt-6-luna13 of 30 - Claude
claude-haiku-5-511 of 30 - Gemini
gemini-3.5-flash-lite9 of 30 - Perplexity
sonar8 of 30
ChatGPTrun 1 of 2
“Which bookkeeping apps would you recommend for a freelancer?”
For most solo workers I recommend Sample Co: invoices and tax set-aside sit together. Rival A is broader, Rival B is cheaper.
- Named
- yes
- Recommended
- yes
- Cited
- 1 of 5 sources
- Price said
- none
Every number opens the answers behind it. Open the full sample report
Our own sites, measured
What they taught usForthsay’s rules came from watching the founder’s own sites in ChatGPT. Same founder, real records: their own analytics, their server logs and our check of each one. Each site has a page with what a plan would track for it.
79%
of first visits came from chatgpt.com, 2 Oct to 31 Oct.
- 79%chatgpt.com 2,146
- 17%direct 474
- 4%google.com 71 · bing.com 15 · other 26
- 12 OctSet up
- 14 OctFirst ChatGPT reads
- 15 OctWhole-site crawl
- 18–25 Oct632 live reads
0 of 32 answers named it when we asked the same four engines through their APIs on 7 Nov. Generic questions miss how its ChatGPT visitors find it.
What a plan would track for CuescribeA new site with no links in. Does it get found?
- 0pages read by OpenAI’s search crawlerto 25 Oct; it fetched robots.txt every day
- 17pages read by Bing’s crawlerto 25 Oct; none indexed yet
- 74requests faking an OpenAI botin one week, from IPs OpenAI does not list
Free maths practice and worksheets for children. Search consoles: Google verified; no public sign of Bing (read 7 Nov). Links from our other sites: none, on purpose.
pencilhop.com is shown without a link on purpose: links from our other sites would spoil the test.
Four parts, one record
01MeasureYour questions, asked of four engines 2 times a week. Named, recommended and cited, each with a count, a 95% range and the answers behind it.
02CheckPrices, dates and claims in the answers, set against your own facts file, with the sentence and the source the engine used.
03SayA dated, sourced facts page for people and the same facts for machines. One truth, readable by both.
04ProveMark a change, and see if anything moved after it. “No change you can trust” is a real answer here.
Free tools, no sign-up
All 13Free brand checkAsk four AI engines about you, twice each, and read all 32 answers.Bot viewFetch a page as a browser and as AI crawlers. See if they get the same thing.Raw HTML readerRead a page the way a reader without JavaScript gets it.Robots readerSee which AI crawlers your robots.txt lets in, by what each one does.Search setup checkIs your site in Google Search Console and Bing Webmaster Tools? What to do if not.Facts file builderWrite dated, sourced facts about your business, for people and for machines.First-screen previewSee your title and the first 200 characters after the headline.Answer graderPaste any AI answer. See if you were named, recommended or cited.
How we measure, next to a common approach
Why it mattersA common approachWhat Forthsay does
One score out of 100Three counts you can redo by hand: named, recommended, cited, each with its denominator and a 95% range.
A rank position inside an answerWhether you were a pick at all. Lists rarely repeat, so a position moves by chance.
Driving logged-in chat apps with scriptsOfficial APIs only, in fresh calls with no memory. We say where API answers differ from the app.
Failed answers counted as “not mentioned”Timeouts, refusals and empty answers kept apart as “not measured”, never as a zero.
Pages served only to AI crawlersThe same page for people and bots. A facts file always has a page people can read with the same words.
Every weekly wiggle shown as newsBefore and after on the same questions, with the range. Most weeks the honest answer is “no change you can trust”.
These are patterns seen across the field, not claims about any one company. Some of them suit some teams.
Plans
Compare- Starter, $39 a month25 questions on 4 engines, about 866 answers a month. For one site you run yourself.
- Growth, $99 a month75 questions on 4 engines, about 2,598 answers a month. For a team that ships changes and wants to see if they worked.
- Scale, $249 a month200 questions on 4 engines, about 6,928 answers a month. For agencies and several brands.
First payment refunded in full within 14 days. Cancel any time.
Let Claude watch it for you
API and MCPAdd Forthsay as an MCP server and ask your assistant “did last week’s pricing page change move anything?”. It reads the numbers and the receipts, not a guess. Six tools work with no account.
{ "mcpServers": { "forthsay": {
"type": "http", "url": "https://forthsay.com/mcp" } } }Paste it into Claude and it adds the server itself, then tries a tool.