How Forthsay measures. Plain rules, and why each exists.

Each number in a report can be redone by hand from the answers behind it. The rules came from watching our own sites in ChatGPT; the records are below.

What counts

Plain rules, not a model. Every result shows the sentence that decided it.

NamedYour brand, one of your aliases, or your domain as a whole word, any case. “Pine” does not match “Pineapple”.

For most freelancers I recommend Sample Co: invoices and tax set-aside sit together.

Named: yes

RecommendedA sentence names you and recommends (recommend, suggest, top pick, best for, try, worth considering, or a pick line such as “Easiest: you”) with no warning in it. Or you open a list of two or more picks.

For most freelancers I recommend Sample Co: invoices and tax set-aside sit together.

Recommended: yes

CitedYour domain, or a subdomain of it, is among the sources the engine read. Look-alike domains do not count.

Sources: sample.co/pricing · freelanceguide.example · rival-a.example

Cited: 1 of 3

ClaimsIn sentences that name you: a price not in your facts file, or a founded year that differs from it. A prompt to read, never a verdict.

Sample Co costs $29 a month and adds tax reminders.

Your facts file: $24 a month

Our own sites, measured

Makerfern went from about 42 ChatGPT visitors a day to 132.

Makerfern is a free PNG to SVG converter for Cricut and laser makers, by Forthsay’s founder. Each bar is one day of visitors who came from chatgpt.com, from its own analytics.

22–28 Sep: 42a day
26–29 Oct: 132a day
2–5 Nov: 115a day

Point at a day to read it. Shaded days had a known fault.

24 SepIndexNow sent all 70 pages to Bing and the other engines that take it.
4 OctA page and a tools list written for AI assistants went live.
9 Oct–15 OctChatGPT’s links lost their tag for a week (a GPT-6 launch bug). Visits showed up as “direct”, so this dip is not a real drop.
30 Oct–1 NovVisitors fell by about half on two of our sites at once, the day after OpenAI’s 29 Oct GPT-6 update. Back to about 90% from 2 Nov.
18–25 OctNot read yet. We leave the gap rather than guess.

Launched 19 Sep with Google Search Console and a 26-page sitemap. Not yet in Bing Webmaster Tools on 9 Oct, so ChatGPT was sending visitors before Bing was set up. These are records of what happened alongside each step, not proof of what caused it.

Why each rule exists

Forthsay, Makerfern, Cuescribe and Pencilhop have the same founder. Numbers come from those sites’ own analytics and server logs, with bots checked against each vendor’s IP list.

Not measured is not absent

Makerfern, 9–15 Oct

ChatGPT visitors seemed to fall by two thirds. They had not: a launch bug stripped the tag from ChatGPT’s links, so the visits were counted as “direct”.

So a timeout, a refusal or an empty answer is “not measured”, shown apart and left out of every share. An outage never shows as a drop.

“No change you can trust” is an answer

Makerfern and Cuescribe, 30 Oct

Both sites lost about half their ChatGPT visitors on the same day, the day after OpenAI updated GPT-6. A change marked that week would have looked like a failure.

So every before-and-after has a range, uses only questions asked on both sides, and says what moved, not why.

Check the claims, not just the mentions

Cuescribe, 16–25 Oct

The price changed on 16 Oct, but one hand-typed page description kept the old $9 a month until 20 Oct. On 25 Oct ChatGPT still quoted $9 from a search copy no crawler had refreshed.

So the claim audit sets every price and year in an answer against your facts file, with the sentence and the source the engine used.

Get into both search consoles

Cuescribe, 12–15 Oct

Google Search Console and Bing Webmaster Tools on 12 Oct; ChatGPT opening pages for people from 14 Oct; OpenAI’s whole-site crawl on 15 Oct. Makerfern got ChatGPT visitors before Bing was set up, so it is a head start, not a rule.

So every report checks for public signs of both, and the free search setup check says what to do if they are missing.

Only count crawlers you can prove

Pencilhop, 19–25 Oct

In its first week, 74 requests used an OpenAI bot’s name from addresses OpenAI does not list. OpenAI’s real search crawler had read robots.txt daily and not one page.

So the crawler log verifier checks each visit against the vendor’s published IP list before it counts.

Ask with web search on, every time

Cuescribe, 18–25 Oct

632 pages were opened inside ChatGPT chats in one week, 55 to 98 a day. Answers come from what an engine reads now, not only from what it learnt.

So every question goes out with each vendor’s web search tool on, and we keep the sources it read.

What we ask

You choose the questions, written the way a person asks them. We sort each one: branded if it names you (“Is Sample Co any good?”), unbranded if it does not (“Which bookkeeping app should a freelancer use?”). The two are never blended into one share, because branded questions name you by design. The headline is your recommendation share on unbranded questions.

How we ask

Through each vendor’s official API, with its web search tool on, in a fresh call with no memory, no account history and no system prompt, 2 times a week. Every answer is stored with the engine, the exact model, the time, the sources it read and the searches it ran. We never script, log in to or copy answers from the ChatGPT, Claude, Gemini or Perplexity apps.

What API answers cannot tell you

Apps add memory, location, personal settings and their own search. One vendor-run comparison found only 24% overlap in brands named between API and app answers for the same questions (Surfer, 2026). So treat API shares as a steady benchmark. To compare, paste answers from your own chats into the app: they are graded the same way and shown beside the numbers, never mixed in.

Ranges

Every share has a 95% Wilson range. Named in 41 of 120 answers is 34%, likely between 26% and 43%. With few answers the range is wide, and the report says so.

ChatGPT13 of 30

Claude11 of 30

Gemini9 of 30

Perplexity8 of 30

Sample data: recommended, one week, 30 answers an engine.

Did the change work?

For each change marker we take four weeks before and four after, on questions asked on both sides only, and compare with Newcombe’s interval for a difference. The verdict is “up” or “down” only when the whole range sits on one side of zero; otherwise it is “no change you can trust”. The free change checker uses the same sums.

Question changes are logged

Adding or removing a question changes what a share is made of. Every change to questions and rivals is logged with its date and shown in the report.

Changes to this page

2026-11-06 First version of the rules below.
2026-11-07 Added what our own sites taught us, with the records behind each rule. No rule changed.