How does ChatGPT choose sources?
ChatGPT answers from two places: what it absorbed during training, and pages it fetches live when it decides to browse. Answers carrying a source card almost always come from the second route, and that route runs on a search index. If a search engine cannot find your page, ChatGPT will not cite it.
The browsing path is worth understanding step by step, because every tactic later in this guide attaches to one of these four stages.
- 01
It reads the prompt and decides whether to browse.
Simple, stable, well-known facts get answered from training data with no sources attached. Anything current, local, commercial or contested triggers a search.
- 02
It runs its own searches, not yours.
The model rewrites your prompt into several machine-friendly queries and fires them at a search index. Those queries rarely match the phrasing you typed, which is why a page can rank for your exact keyword and still never be seen.
- 03
It fetches and skims the candidate pages.
This is where structure decides everything. The model is looking for a passage that answers the rewritten query cleanly enough to quote. Pages that bury the answer under six paragraphs of preamble lose to pages that lead with it.
- 04
It composes the answer and attaches the source cards.
Usually three to eight sources survive. Being retrieved is not the same as being cited, and the gap between the two is almost always structural.
Three separate crawlers sit behind this. One gathers training data, one builds the search index that browsing queries hit, and one fetches a specific page in real time when a user asks about it. They are all declared in robots.txt and they are all blockable. Blocking the training crawler while allowing the search crawler is a legitimate choice. Blocking all three removes you from the answer set entirely, and we have seen sites do this by accident when a security plugin tightened robots.txt without anyone reading it.
Is ChatGPT SEO the same as ranking in Google?
No, but the overlap is large enough that treating them as separate budgets is a mistake. Classic SEO is the substrate. ChatGPT citation is what you build on top of it.
Two published findings frame the debate. Semrush analysed one brand's citations and reported an 85 per cent correlation between ChatGPT citations and Google keyword rankings. seo.com reported that ChatGPT Search results are around 73 per cent similar to Bing's. Both are real, both are partial, and nobody reconciles them.
Our reading is that they measure two different links in the same chain. Bing similarity explains retrieval: the index decides which pages are eligible. Google correlation explains selection: the qualities that make Google rank you, namely relevance, authority and clean structure, are the same qualities that make a model quote you. Neither finding says Google rankings cause ChatGPT citations. They say the same underlying work produces both.
The practical consequence is that you do not need a separate ChatGPT strategy. You need your existing SEO plus four additions: Bing indexation, extractable structure, entity consistency off-site, and measurement that can see AI referrals. That is the whole job.
| Finding | What it measures |
|---|---|
| Bing similarity ~73% (seo.com) | Retrieval: which pages are eligible |
| Google ranking correlation 85% (Semrush) | Selection: which eligible page gets quoted |
What makes a page citable by ChatGPT?
A citable page answers one question directly, near the top, in language a model can lift without editing. Everything else is a variation on that.
01
A definitive opening answer.
Forty to sixty words, immediately under the heading, no throat-clearing. This single change does more than any other.
02
Question-shaped headings.
Write headings as the question a person would type or say, then answer it in the first sentence beneath.
03
One idea per passage.
Models retrieve passages, not pages. A section that covers three things is three half-answers.
04
Tables and short lists.
Structured comparisons get quoted disproportionately, because they survive being pulled out of context.
05
Numbers with sources attached.
A claim with a named source and a date is safer for a model to repeat than a claim without one.
06
Something only you can publish.
Original data, a tested method, a real result. Models prefer to quote the origin of a fact rather than the fourth site to repeat it.
07
Visible dates.
Publish date and last-reviewed date, in the page, not just in the schema. Currency is a selection signal.
08
Clean, crawlable HTML.
Server-rendered text, real headings, real tables. Content that only exists after a client-side render is content that may never be read.
How to rank in ChatGPT: nine steps that earn citations
Work through these in order. The first three are prerequisites; the rest compound.
01
Get indexed in Bing.
Verify the site in Bing Webmaster Tools, submit the sitemap, fix whatever it reports, and enable IndexNow so changes are picked up quickly. This is the single most repeated tactic in the field and the one most often skipped. A page Bing has never crawled cannot be retrieved.
02
Keep ranking in Google.
The correlation is high and the work is shared. Nothing in this guide replaces technical health, useful content and links.
03
Let the right crawlers in.
Read your robots.txt. Decide deliberately which of the three crawlers you allow, and check that a firewall or bot-protection rule is not quietly overriding your decision.
04
Restructure your top twenty pages for extraction.
Definitive opener under each heading, question-shaped H2s, one comparison table, short paragraphs. Do the pages that already earn commercial traffic first.
05
Ship the schema.
Organization on the site, Article on posts, FAQPage on genuine question blocks, and sameAs pointing at every profile you control. Schema does not force a citation. It confirms what the page says and who published it.
06
Fix the About page and your entity data.
Legal name, trading name, address, founders, services, and the same details repeated identically everywhere they appear. Inconsistent entity data is how a model ends up unsure whether two mentions are the same company.
07
Earn mentions where models already look.
Industry listicles, comparison sites, review platforms, relevant subreddits, trade press. Unlinked mentions count here, which is not true of classic link building.
08
Publish original data.
A survey, a benchmark, an analysis of your own account base. This is the one step your competitors will not copy, because it costs real time.
09
Add dates and keep them true.
Show the publish date, show the review date, and actually review on that cadence. A stale date is worse than none.
How do I know if ChatGPT has mentioned my brand?
Ask it, properly and repeatedly. A one-off prompt from your own logged-in account proves nothing, because memory and past conversations skew the answer. Run a fixed prompt set, in a signed-out or temporary chat, with memory off, once a month, and record what comes back.
Below is the exact set we run. Replace the bracketed terms and keep the wording otherwise identical, so month-on-month results stay comparable. Record three things per prompt: were you mentioned, were you cited with a link, and which competitors appeared.
The monthly visibility prompt set
- 01Best [service] agencies in [city or country]
- 02Who are the leading [service] specialists for [industry] businesses?
- 03I need help with [problem]. Which companies should I shortlist?
- 04What does [your brand] do?
- 05Is [your brand] any good? What do people say about them?
- 06Compare [your brand] and [competitor]
- 07How much does [service] cost in the UK?
- 08What should I look for when choosing a [service] provider?
- 09[Your brand] reviews
- 10Who publishes original research on [your topic]?
Prompts four, five and nine test whether the model knows you exist and describes you correctly. Prompts one, two, three and six test whether you are in the consideration set, which is the commercially valuable position. Prompts seven, eight and ten test whether your content is being used as reference material, which is where citations come from. A brand can score well on the first group and badly on the other two, and that pattern tells you to fix content structure rather than brand awareness.
Score it simply. Mentions divided by prompts run gives you a visibility rate. Track that one number monthly and you will see movement long before it shows up in traffic.
How to get cited in AI search: corroboration and consistency
Models trust claims that more than one independent source agrees on. Your own site states what you do; third-party sources confirm it. Citation follows agreement.
That is why off-site work matters more in AI search than it did in classic SEO, and why unlinked mentions suddenly count. Three source types do most of the lifting.
Aggregators
Aggregators and listicles.
Best-of roundups, directories and curated lists in your sector. These are heavily retrieved for shortlist prompts, which are the prompts with buying intent behind them.
Community
Community and review platforms.
Forum threads, subreddits and review sites carry weight because they read as independent testimony. You cannot manufacture these credibly, and attempting to is a fast route to being described badly.
Trade press
Trade press and named commentary.
Being quoted as a named expert attaches a person to a company to a topic, which is exactly the association a model needs to make.
Consistency is the multiplier. Same legal name, same trading name, same address, same service names, same founder name and title, everywhere. If your company is described three different ways across your site, your listings and your profiles, you have given the model three weak signals rather than one strong one. We fix this before anything else on client accounts, because it is cheap and it unblocks everything downstream.
How to track brand ranking in ChatGPT
2
Enquiries traced to ChatGPT, July 2026
1
GA4 channel: AI Assistant
10
Prompts in our monthly panel
AIO
Cited on multiple commercial terms
Set up an AI Assistant channel in GA4, then match it against your CRM. Traffic from ChatGPT arrives as a referral, so it is visible without any special tooling, but by default it lands in the referral bucket and disappears into the noise.
Do this once and it reports forever.
- 01
Create a custom channel group in GA4 called AI Assistant, matched on referral source containing chatgpt, perplexity, claude, copilot and gemini domains. Keep the definition in a document, because you will want the same one across accounts.
- 02
Add the same segment in Looker Studio or your reporting layer, so AI sessions appear next to organic rather than buried inside referral.
- 03
Log the landing page. AI referrals concentrate on a handful of pages, and those pages tell you exactly what is being cited.
- 04
Pass the channel into your enquiry form as a hidden field, so the attribution survives into the CRM. Without this you can see sessions but never revenue.
- 05
Cross-check branded search in Search Console. A rise in branded queries with no matching campaign usually means AI answers are doing the introduction and search is doing the click.
Two warnings. Referral data undercounts, because some AI traffic arrives with no referrer and lands in direct. And the manual prompt panel from the previous section measures something the analytics cannot see at all: the answers where you are mentioned but nobody clicked. Run both.
What two ChatGPT enquiries actually looked like
In July 2026 we received two new business enquiries that we could trace to ChatGPT end to end. Small numbers, and we are presenting them as a case rather than a benchmark.
The chain was the same for both. The AI Assistant channel in GA4 recorded the session with a chatgpt.com referrer. The landing page in each case was a page we already knew was being cited in Google AI Overviews on commercial terms, not the home page. The visitor read two to three pages, completed the standard enquiry form, and the hidden channel field carried AI Assistant through into the CRM record, where it is still attached to the deal.
Two things stand out. First, the entry pages were structured pages, the ones with a direct opening answer and a comparison table, not our longest essays. Second, the same pages were already earning AI Overview citations, which supports the argument that the underlying work is shared rather than split. The full referral dataset behind our AI channel reporting is published separately.
How can I use ChatGPT for SEO optimisations?
Use it for volume work with a verifiable output, and never for facts. That is the honest boundary, and it has not moved much in two years.
Genuinely useful for
- Drafting and reshaping. Turning a messy brief into an outline, tightening a paragraph, generating heading variants.
- Clustering and classification. Sorting a keyword export into topics, tagging intent, spotting near duplicates.
- Structured data. Producing valid schema from content you supply, then validating it.
- Bulk repetitive tasks. Alt text, meta description first drafts, internal link suggestions from a list of URLs you provide.
- Adversarial review. Asking it to argue against your own page and list what a sceptical reader would ask.
Do not use it for
- Search volume and difficulty. It does not have them. It will produce plausible numbers anyway.
- Competitor rankings and backlinks. Same problem, more confidently wrong.
- Statistics and citations. It will attribute a real-sounding figure to a real-sounding source that never published it.
- Finished copy at scale. Unedited output is generic by construction, and generic is the opposite of citable.
- Anything you cannot check. If you would not publish it without verifying it, do not publish it.
The short version: it is a fast junior with no memory of the truth. Give it structure to work with and check everything it hands back.
What to do if ChatGPT gets your brand wrong
Fix the source, not the answer. There is no button that edits what a model says about you, so the only durable correction is to change what it reads.
01Find the origin. Ask the model where it got the claim, then check those pages. Usually it is an outdated listing, an old press mention or a competitor comparison.02Correct the third-party sources. Update directory entries, ask publishers to amend, and refresh anything you control that carries the old detail.03Publish an unambiguous statement on your own site. A clear, dated page that states the correct fact plainly gives the model something better to retrieve.04Re-run the prompt set after four to six weeks. Corrections take a cycle to appear, and you need the before and after to know whether it worked.
Common mistakes
Do
- Lead every section with the answer.
- Verify the site in Bing Webmaster Tools this week.
- Publish one thing a year that only you could publish.
- Keep entity details identical everywhere.
- Measure with both analytics and a manual prompt panel.
Don't
- Do not block the crawlers by accident.
- Do not add an FAQ block that repeats the article instead of answering new questions.
- Do not chase a citation with thin content and heavy schema. Schema describes; it does not persuade.
- Do not treat this as separate from SEO with a separate budget.
- Do not judge visibility from one prompt in your own logged-in account.
ChatGPT SEO: frequently asked questions
Methodology
Sources: our own GA4 AI Assistant channel data and CRM records for the July 2026 enquiries; our monthly ten-prompt visibility panel, run signed out with memory off; published findings from Semrush and seo.com, attributed in the text. First-party AI referral datasets are published in full in the linked statistics posts and are not restated here. Last reviewed: August 2026. Next review: February 2027.
Work With Visionary Marketing
Want to know whether ChatGPT cites you?
We run the prompt panel, the Bing and structure work, and the GA4 tracking. Free 30-minute audit, no obligation.
Visionary Marketing is a UK-based SEO and Google Ads agency that takes a data-led approach to growth. We don't guess - we analyse your market, competitors, and performance data to build strategies that drive measurable revenue. Every campaign is grounded in real numbers, not assumptions.
