SEO for ChatGPT: What Actually Moves the Needle

By Muntazir · Published 10 September 2026

Ask ChatGPT for the best accountant in Manchester and it doesn’t return ten blue links. It writes three sentences and names two firms inside them. There’s no page two to fall back to — the selection already happened, invisibly, before the answer appeared.

That’s the shape SEO for ChatGPT actually has to deal with: not a new set of tricks bolted onto the old ranking game, but a different kind of selection, built from pieces that don’t update on the same schedule.

What is SEO for ChatGPT?

SEO for ChatGPT means structuring content so the model’s retrieval and reasoning steps can find a page, trust it, and lift a fact out of it cleanly — rather than optimizing for an algorithm that scores keyword relevance across ten results. Someone typing “best accountant in Manchester” into ChatGPT gets a short list embedded in prose, built from whichever sources the model both located and judged reliable enough to name. Missing that list costs a lead outright, the same way a page eleven costs one on Google, except there’s no way to scroll to page eleven here. The broader mechanics of how a model reads a page differ from a search engine’s in ways that matter before any of the rest of this does.

Does ChatGPT use the same ranking signals as Google?

No, and the difference starts with what “ChatGPT” is actually doing when it answers. OpenAI’s own crawler documentation describes three separate crawlers with three separate jobs. GPTBot collects pages that may end up in a future training run — it feeds the model’s frozen, parametric memory, the kind that only changes when a new version is trained. OAI-SearchBot builds the index ChatGPT’s search feature queries in real time, closer to how a conventional search engine’s own crawler works. ChatGPT-User fetches a page on the spot when someone asks the model to open a specific link — a proxy for a person clicking, not a crawl at all.

A site can allow one of these and block another, and the two decisions do different things. Block GPTBot and a business stops being learned by future models, while whatever the current model already absorbed stays untouched. Block OAI-SearchBot and the business disappears from what ChatGPT can currently look up and cite, even if the model still recognizes the name from training. Google runs one signal set through one crawler. ChatGPT runs two separate systems on two separate schedules, and a page can be well positioned in one while invisible to the other.

What content structure does ChatGPT prefer?

Researchers at Princeton University and IIT Delhi tested this directly rather than guessing at it. Their 2024 paper on generative engine optimization measured which page-level changes moved a source higher in AI-generated answers, and found that adding citations, direct quotations and statistics to source text improved visibility by roughly 30 to 40 percent — more than any wording or keyword-density change they tried (arXiv:2311.09735). Improving fluency and readability helped too, but by a smaller margin, and keyword stuffing wasn’t among the strategies that worked at all.

That matches what the citation-mechanics side of this already covers at the sentence level: short, self-contained passages that state a fact plainly are what a model can lift out of a page and attribute correctly. A page built from vague adjectives and long, dependent clauses gives it nothing clean to quote, no matter how well that page ranks anywhere else.

Does ChatGPT browsing use real-time SEO signals or training data?

Both, and which one answers a given question depends on the question. Something ChatGPT can answer instantly from memory — “what is a canonical tag” — draws on the frozen training snapshot, gathered by GPTBot and locked in whenever the model was last trained. Fix a page today and that particular answer won’t change until a future model is trained on the fixed version, on a timeline OpenAI doesn’t publish.

A query that triggers ChatGPT’s search tool — often anything time-sensitive, local, or phrased like a live lookup — instead hits the OAI-SearchBot index, the one that actually powers live search results rather than training. OpenAI doesn’t publish how often that index itself refreshes; what its documentation does confirm is that a robots.txt change takes about a day to take effect on it, versus a model retrain that happens on a release cycle measured in months, not days. The exact cadence isn’t public, but the two processes are structurally different speeds, and a page fix reaching a live index is the faster of the two almost by definition.

That split explains a complaint that otherwise looks like a bug: a business fixes its site and ChatGPT keeps describing it the old way. The fix may already be sitting in the fast-moving search index while the outdated description is coming from the slow-moving trained memory — two different stores of “what ChatGPT knows,” repaired at two different speeds, and there’s no setting that forces the second one to catch up.

How do you check if ChatGPT already knows your business?

Two different questions, and they need two different tests. The first is recall: asked “what is [business]” with no other context, does a model know anything real about it, or does it hedge and describe what the words in the name usually mean instead of the company? The second is recommendation: asked “best [category] in [location]” without the business named at all, does a model volunteer it unprompted? A business can score well on the first and never appear in the second — recall and recommendation are separate outcomes, and only one of them produces a customer.

Both tests are expensive to run properly: several models, several phrasings of each question, several samples of each phrasing, because a single greedy answer hides the run-to-run variance that’s part of what’s actually being measured. That cost is exactly why a free scan of a site doesn’t attempt either one. It runs a single, inexpensive model pass that reads the page itself and reports back a comprehension snapshot — how clear the offering is, what category a customer would search under — without asking any model what it already believes about the brand. Recall and recommendation testing, sampled repeatedly across several models rather than asked once and trusted, sit behind the paid deep scan.

A site optimized for both crawlers and structured for citation still can’t force a correction into a model that’s already trained wrong about it. It can only make sure the next version, whenever that arrives, has something better to learn from.