What ChatGPT Actually Looks for When Recommending Experts
Forget backlinks and authority scores. ChatGPT recommends experts using 6 specific signals most coaching sites fail. Here's what each one looks like in practice.
Built BakingSubs to 162,500 Copilot citations and accelerating. Now teaching the system behind it.
- ai-visibility
- chatgpt
- strategy
- case-study
Ask ChatGPT to recommend a life coach for new mothers in London, and it won't pick the site with the most backlinks. It will pick the one whose page answers that exact question in a way ChatGPT can lift cleanly. That's a different game than SEO, and most expert sites are still playing the old one.
Key takeaways
- ChatGPT does not rank pages by authority or backlinks. It picks sources whose answers are easy to quote and clearly tied to a specific question.
- The six signals that matter are: easy-to-quote answers, niche specificity, first-party data, clear author identity, structured layout, and topical coverage.
- A small site with 30 deeply specific posts gets cited more often than a large site with 300 broad ones. This is why BakingSubs earned 144,321 Microsoft Copilot citations in a single quarter.
- The most overlooked signal is unique first-party data. Anything you have measured, tracked, or seen with your own clients that nobody else has published is gold.
- You can check how visible you are right now with the free AI Visibility Check, which runs 8 buyer-style questions through each engine.
Why ChatGPT recommendations don't work like Google rankings
Google was built to rank pages. ChatGPT was built to answer questions. That sounds like the same thing but it isn't, and the difference is what trips up most experts.
When Google ranks a page, it asks: is this site authoritative, well-linked, and relevant to the search term? When ChatGPT picks a source to cite, it asks something closer to: can I lift a clean, specific answer from this page and stand behind it? Authority helps a little. Backlinks help less than people think. What matters most is whether the page contains a sentence ChatGPT can paste into its answer without ambiguity.
This is why huge sites with thousands of links often lose to small expert sites. A coaching directory page that lists 40 coaches in vague language is harder for ChatGPT to quote than a single coach's page that says "I work with women in their first year of postpartum recovery, three months minimum, $2,200." One is a list. The other is an answer.
If you want to go deeper on the mechanics, I broke down how coaches show up in ChatGPT search and what the engine is actually doing under the hood.
Signal 1: Answers ChatGPT can lift in one sentence
The first thing ChatGPT looks for is whether your page contains a clear, self-contained sentence that answers a specific question. Not a paragraph. Not a section. A sentence.
Most expert sites bury their answers inside long, hedged paragraphs. "There are many factors that go into deciding whether group coaching is right for you, and it really depends on your situation, but generally speaking…" ChatGPT cannot use that. It needs a sentence like: "Group coaching works best for founders who already have a paying product and need accountability, not for people still figuring out what to build."
Look at any post on your site right now. Pick a buyer question it should answer. Search the page for a single sentence that answers it cleanly. If you can't find one, ChatGPT can't either.
Signal 2: Niche specificity that excludes the wrong people
The second signal is how narrowly you describe who you work with. Counter to common advice, ChatGPT prefers sources that exclude readers, because exclusion proves the answer fits the specific buyer.
A site that says "I help busy professionals find balance" gets passed over. A site that says "I help corporate lawyers in their fourth to seventh year reduce 60-hour weeks without leaving the partner track" gets cited. When a buyer asks ChatGPT "who helps senior lawyers avoid burnout without quitting," the second site is a clean match. The first one matches nothing in particular.
Priya, a life coach in Toronto, spent six months trying to be findable for "life coaching." Nothing happened. She rewrote her homepage and three posts to focus on second-generation South Asian women in finance navigating career and family pressure. Her first ChatGPT citation came seven weeks later, on a query she hadn't even targeted directly: "coach for high-achieving immigrant daughters." The specificity was the unlock. The same pattern shows up for coaches narrowing their niche to fill a pipeline and for consultants who keep losing to vaguer competitors.
Signal 3: First-party data nobody else has
The third signal, and the most underused, is unique first-party data. ChatGPT is hungry for numbers, observations, and patterns it can't find on twenty other sites.
First-party data is anything you have seen, measured, or tracked with your own clients or your own work that nobody else has published. It does not have to be a formal study. It can be:
- "Across my last 40 clients, the ones who showed up to every weekly session in month one were the ones still active in month six. Skipping the second session was the strongest predictor of dropping out."
- "Workplace mediations I've run between founders end in agreement 80% of the time. Mediations between cofounder and investor end in agreement closer to 20%. The dynamic is completely different."
- "Personal training clients over 50 who lift three times a week recover better than the same group lifting four times. I switched my entire over-50 cohort to three-day programs last year."
These are the kinds of sentences ChatGPT will quote, because they exist nowhere else. They also signal you have done the work. BakingSubs got cited 144,321 times in a single quarter partly because every substitution post included data from actual baking tests, not just summaries of what other sites said. That is the engine's favorite kind of source.
If you have never written down what you have noticed across your own clients, that is the single highest-leverage thing you can do this month.
Signal 4: Clear author identity tied to real expertise
The fourth signal is whether ChatGPT can tell who wrote the page and why they would know.
This is where Claude and ChatGPT diverge a little. Claude weighs author signals more heavily, but ChatGPT still wants to see a real human attached to the claims. A page with no author, no bio, and no first-person experience reads to the engine like a generic article. A page with a named author, a short credibility line, and first-person language ("I've worked with 60 founders over the last four years…") reads like a source.
The minimum bar:
- A real name on every post, not "Admin" or "Team"
- A short bio sentence at the top of long posts: who you are, what you do, why you would know this
- A Person schema block on your About page (the hidden tag that tells AI engines this page is about a real human, not a brand)
- First-person language in the body, not just objective third-person summaries
Most expert sites I look at fail at least two of these. The fix takes an afternoon.
Signal 5: Structured layout the engine can parse
The fifth signal is whether your page is easy to scan. Not just for humans. For the engine, too.
ChatGPT reads HTML structure. A page with one clear H1, descriptive H2s, an FAQ section at the bottom, and a clean question-then-answer pattern is easier to extract from than a wall of unbroken prose. This is not about gaming structure. It is about removing ambiguity.
A plausible failure case: a coach I looked at had a long About page with 14 H1 tags scattered through it, each labeling a different section ("My Story," "My Approach," "Who I Work With"). Claude treated the page like a category index rather than a recommendation candidate, because category pages have many H1s. Once she collapsed them to one H1 and proper H2s, the page started showing up in test queries within two weeks.
Keep it simple. One H1 per page. Descriptive H2s in sentence case. A FAQ section at the bottom of any post that answers buyer questions. Don't overthink it past that.
Signal 6: Topical coverage across a cluster, not one-off posts
The sixth signal is whether you have written enough about one specific topic that the engine sees you as a go-to source on it.
This is the heart of the Citation Cluster Method. One post on a topic, no matter how good, rarely gets cited. A cluster of 8 to 15 posts that all answer related questions on the same narrow topic gets cited again and again, because every time a buyer asks something in that topic area, your site is the one with a clean answer.
BakingSubs did this with baking substitutions. Every common ingredient swap got its own post, with its own data, its own test results, its own question-and-answer structure. After 12 months, it had earned 5,000+ daily Google clicks and the 144,321 Copilot citation count, with no ads, no backlinks campaign, no social media.
For a coach or consultant, this means picking one narrow problem your buyers face and writing 10 posts that answer the 10 questions they ask about it. Not 10 posts on 10 different topics. 10 posts on the same topic, each answering a slightly different question. That is what ChatGPT learns to recognize you for. I covered the structure of clusters that actually get cited separately, but the principle is: depth in one place beats breadth everywhere.
How the 6 signals stack
No single signal is enough. The sites that get cited heavily hit four or five of these at once, and the strongest ones hit all six.
| Signal | What it looks like | Easiest first fix |
|---|---|---|
| Easy-to-quote answers | One sentence per question, no hedging | Pick 5 buyer questions, write a clean one-sentence answer for each |
| Niche specificity | Excludes the wrong buyer in plain language | Rewrite your homepage hero to name who you don't serve |
| First-party data | Numbers and patterns from your own work | Write down 10 things you've noticed across your clients |
| Author identity | Real name, bio, Person schema | Add a one-sentence bio to the top of long posts |
| Structured layout | One H1, descriptive H2s, FAQ at bottom | Audit your About page for stray H1 tags |
| Topical coverage | 8 to 15 posts on one narrow topic | Pick one topic and commit to a cluster, not scattered posts |
Most coaching and consulting sites I see hit one or two of these by accident. They got specific enough on their homepage, or they happen to have a clear bio. The ones that win on AI search hit all six on purpose.
Frequently asked questions
How long does it take to get cited by ChatGPT after I fix these signals?
In my experience, the first citations on test queries usually show up four to eight weeks after the changes go live, depending on how often the engine refreshes its sources. Compounding kicks in around month three, when multiple posts in a cluster start getting picked up together. BakingSubs took 12 months to reach the 144,321 number, but the first citations appeared within the first two months.
Do backlinks matter at all for ChatGPT recommendations?
A little, but far less than for Google. Backlinks help the engine trust that you are a real source, but they will not overcome a page that lacks specificity or first-party data. I would rather have one strong post with unique data and no backlinks than ten generic posts with strong backlinks. If you want the broader picture on what's changed, I wrote about what replaces SEO when buyers stop Googling.
What if I'm in a regulated niche like health coaching?
Health, finance, and legal niches face a higher bar because AI engines weight credibility signals harder on topics that affect a buyer's wellbeing. The six signals still apply, but author identity and first-party data carry extra weight. Health coaches get recommended by leaning hard on credentials, real client outcomes, and conservative claims.
Should I write for ChatGPT or for human readers?
For human readers, written in a way ChatGPT can extract from. The two goals do not conflict. Clear, specific, well-structured writing serves both. The mistake is writing for ChatGPT in a way that reads robotic to humans. Specific, plain-language answers win on both sides.
Can a brand-new site get cited, or do I need years of history?
A new site can absolutely get cited, but it needs to hit the six signals harder than an older site would. BakingSubs was built from scratch in 12 months. The advantage of a new site is you can architect everything around these signals from day one, instead of retrofitting an older site that was built for Google.
What to do this week
Pick one narrow buyer question you should be the answer to. Write down a clean one-sentence answer to it. Look at the page on your site that should hold that answer. If the sentence is not already there, paste it in.
Then do the same thing four more times. Five questions, five sentences. That is enough to move the needle on the first signal, which is the foundation for all the others. When you want to see exactly which engines are already finding you and which ones aren't, run the AI Visibility Check and you'll get the eight-question diagnostic for each one.