Surfaces our work more for you across Google Search and AI answers.
Open your best service page and read the first three sentences out loud. If they welcome the visitor, mention your passion for quality, and promise a family-owned experience, they are the sentences ChatGPT is most likely to pick when it decides whether to mention you. Nothing in them tells it what you do, for whom, or at what price.
Quick read: On November 20, 2026, Peec AI, a company that sells AI-search visibility tracking, published an analysis by Tomek Rudzki of what it describes as leaked ChatGPT search data: the system prompt, the searches ChatGPT ran, the pages it retrieved, the score each page received, and the text passed to the model (Peec AI). Peec reports that ChatGPT cuts each page into pieces of about 1,000 characters and judges the page by its best piece; that "in 92% of cases, the selected piece starts in the first 10% of the page"; and that "even when a later piece matched the question better, ChatGPT still picked the first one 56% of the time." In controlled tests that changed one fact on one page, Peec found that a wrong price for a brand known for low prices cost 100% of that page's citations, and that plain praise of a brand got it named in 0% of answers while a description matching the question's role got it named in 100%. Peec says it tried to verify every claim, that "some things here can't be fully checked," and that findings already differ between model versions. The figures are Peec's and have not been independently replicated.
What Peec says it saw
When ChatGPT searches the web to answer a question, it doesn't read your page the way a customer scrolls it. Per Peec's write-up, it breaks each page into pieces of roughly 1,000 characters, about 150 to 200 words. A reranker scores the pieces against the searches ChatGPT ran, and the page's score is the score of its best piece. Around 14 pages make it into the prompt the model writes its answer from.
The position finding is the one to sit with. Peec compared 1,640 pages that GPT-6 used in October 2026 against the live code of those pages. In 92% of cases the piece ChatGPT selected started in the first tenth of the page. On the previous model, GPT-5.6, the first piece was chosen 75% of the time. And when a later piece was the better match, the first one still won 56% of the time.
Three more results from the same write-up matter to a small business:
- One wrong fact can cost the whole citation. Peec changed one fact on one page at a time and ran 40 to 60 answers per version. Calling a budget option "enterprise" cut citations by 44% to 95%. A wrong price for a brand known for low prices cut them by 100%. A wrong price for a brand not known for price made no difference. Peec's summary: "Positioning is the costliest mistake."
- Praise does nothing; a matching role does everything. For a real tool the model knew but didn't name unprompted, a page offering plain praise got it named in 0% of answers. A full product description got 35%. A verifiable claim got 43%. A description that matched the role the question asked about got it named in 100%, always first.
- The domain decides close calls. The same page was cited in 20 of 20 answers on Search Engine Land's domain and 0 of 20 on a Blogspot address. Peec adds that "the domain only decides the close calls": a page with the exact fact the question needed still got cited on Blogspot.
Peec also reports that adding a first-person anecdote to a page dropped the brand's mentions from 40% of answers to 10%.
Why the caveats don't make it useless
Here's the thing about vendor research: Peec sells the tool that measures this, and the data came from a leak Peec itself says can't be fully checked. Treat every number above as Peec's claim. Some of these tests ran 20 answers per condition, which is enough to see a direction and too few to trust the second digit.
What makes it worth reading anyway is that the direction matches what the rest of the field already advised without evidence. "Put the answer near the top" has been standard Answer Engine Optimization (AEO) advice for a while, and the skeleton this blog is built on opens every post with a neutral summary for that reason. Peec's data is the first look we've seen at how strong the effect might be. If it holds even loosely, the top of the page does far more of the work than the rest.
The second finding cuts closer for small businesses. Plenty of small businesses have a price, a service list, or a description floating around somewhere that's two years out of date: an old directory listing, a Facebook "About" box, a rate sheet PDF. Peec's tests suggest that a page stating a price or positioning that contradicts what other sources say can lose the citation outright. That is the same problem as AI assistants describing the version of your business you closed down, measured from the other side.
For the wider picture of how search and AI answers now overlap, we've mapped the territory between being ranked and being quoted.
Rewriting the first 1,000 characters
You don't need to rebuild the site. You need to rewrite the opening of the five or so pages a customer would land on: the homepage, each main service page, and the contact or pricing page. I ran a grooming salon for years, so I know an afternoon is a lot to ask. This fits in one.
For each page, take the first 1,000 characters (paste them into any character counter) and check that they answer these, in plain sentences:
- What you do. The service, named the way a customer would type it. "Mobile dog grooming," not "pampering solutions."
- For whom. The customer and the area at the level you're comfortable publishing: "for dogs and cats across the county," "for small dental practices."
- The role you fill. Peec's 100% result came from a description matching the question's role. If people ask for "a groomer who takes anxious dogs," say you are one, in those words, if it's true.
- A fact someone can check. A starting price, a turnaround time, a license, hours. A verifiable claim beat praise in Peec's tests by 43 points.
- No contradiction with your other listings. Whatever price or positioning you state should match your Google Business Profile, your directory listings, and your social bios. If you don't publish prices, make sure no old page or listing is publishing one for you.
A hypothetical, made up for this post, to make it concrete. A service page that opens "Welcome! At Example Grooming we believe every pet deserves to feel special" gives ChatGPT a first piece with no service, no customer, and no fact. "Example Grooming is a mobile dog and cat groomer serving the whole county. Full grooms start at $[your price], appointments run Tuesday to Saturday, and we take senior and anxious dogs" gives it four. The welcome line can move further down the page, after the facts.
Then do the matching pass. List every place your price, service area, and positioning appear online, and fix the old ones at the source. Peec's own recommendation is to fix wrong third-party descriptions where they live, not just on your site.
Edits that won't move the number
- A longer page. If the selected piece usually starts in the first tenth, adding 2,000 words to the bottom adds to the part of the page that rarely gets selected.
- A testimonial wall at the top. The closest thing in Peec's tests, plain praise, got the brand named in 0% of answers. Reviews matter to the human reading the page. Put them below the facts, not in front of them.
- An origin story as the opening. Your story belongs on the About page. On a service page, Peec's anecdote test points the wrong way.
- Chasing a "domain authority" score. Peec says ChatGPT doesn't use third-party scores like Moz Domain Authority (DA) or Ahrefs Domain Rating (DR), and that domain only tips close calls. A clear first paragraph is cheaper than either.
- A contrarian headline to stand out. In Peec's tests, a well-argued contrarian verdict lost 44% to 77% of its citations. Say what you do and leave the provocation for the blog.
- Treating this as a permanent rulebook. Peec's numbers already shifted between GPT-5.6 and GPT-6. The opening paragraph will still matter next year; the percentages may not.
FAQ
How does ChatGPT choose which websites to cite?
According to Peec AI's November 2026 analysis of leaked ChatGPT search data, ChatGPT runs several searches for a question, splits the retrieved pages into pieces of about 1,000 characters, scores the pieces, and judges each page by its best piece. About 14 pages reach the model. Peec found that clear, specific facts, a description matching the question, and consistency with other sources all raised the odds of being cited. These are Peec's findings and have not been independently confirmed.
Does ChatGPT only read the top of a web page?
No, but Peec reports that the top dominates. In its sample of 1,640 pages used by GPT-6, the selected piece started in the first 10% of the page 92% of the time, and the first piece was chosen 56% of the time even when a later one matched the question better.
How long should the opening of a service page be?
Peec recommends putting answers in the first 1,000 characters, roughly 150 to 200 words. That is enough for three or four plain sentences covering what you do, for whom, where, and one checkable fact like a starting price or hours.
Can an outdated price on another website hurt my ChatGPT visibility?
Peec's tests suggest it can. Changing a page so it gave a wrong price for a brand known for low prices cut that page's citations by 100%, and wrong positioning cut them by 44% to 95%. Keeping your price and description consistent across your site, Google Business Profile, and listings removes the contradiction.
Should I put customer testimonials at the top of my page for AI search?
Not at the very top. In Peec's tests, plain praise of a brand got it named in 0% of answers, while a factual description matching the question's role got it named in 100%. Testimonials still help the people reading the page, so place them after the opening facts.
Is the ChatGPT leak data reliable?
Treat it as directional. Peec AI sells AI-visibility software, says some details "can't be fully checked," ran some tests on 20 answers per condition, and notes results already changed between model versions. The broad pattern, that clear facts near the top of a page get used more, agrees with long-standing advice, but the exact percentages should not be quoted as fact.
If you'd rather someone else rewrite those first 1,000 characters and keep your facts matching everywhere they appear, that's the work behind getting found on Google and in AI answers.
Want a back office for your site?
The build is free on any Care plan with a 12-month minimum. Care starts at $199/mo, and you can switch tiers anytime.
Apply for a build slot →RELATED READS