How Does ChatGPT Decide Which Businesses to Recommend?
There is no submission, no dashboard and no ranking. What there is, documented by OpenAI itself, is a set of crawler controls that decide whether you can be used at all. That part is concrete and checkable. The rest is observation.
What Is Actually Happening
When somebody asks ChatGPT for a business recommendation, the reply can be built from two quite different places. Material fetched from the live web. Or material the underlying model absorbed in training. Those behave differently and only one of them is influenced by anything you do this month.
The search route. Documented. The one that matters commercially.
OpenAI's crawler documentation, checked on 30 July 2026, describes a crawler called OAI-SearchBot whose stated purpose is to surface websites in ChatGPT's search features. That is the route by which current information about your business can reach an answer.
The training route. Also documented, though largely outside your control.
The same OpenAI documentation, checked on 30 July 2026, describes a separate crawler called GPTBot used to collect content that may be used in training the foundation models. Anything absorbed that way cannot be edited afterwards, which is why a business whose old description was captured cannot simply correct it.
A third agent, for user actions. Worth knowing about because it is easily confused with the others.
OpenAI's documentation, checked on 30 July 2026, states that ChatGPT-User handles certain actions initiated by a user rather than automatic crawling. The same documentation states plainly that it is not used to determine whether content may appear in Search.
Why the distinction is the most useful thing on this page. It tells you which control to care about.
A great deal of published advice treats all of this as one crawler. It is not. The roles are documented separately, so getting the wrong one wrong is the commonest expensive mistake in this subject.
This page is reviewed monthly. This documentation has changed more than once.
There Is No Submission And No Dashboard
You cannot submit your business to ChatGPT. There is no form, no listing to claim, no verification process and no console showing how you are performing. OpenAI's published material for site owners describes crawler controls rather than any mechanism for putting a business forward.
What that rules out immediately. An entire category of offer.
Anybody selling submission to ChatGPT, registration with it or a guaranteed place in its answers is selling something that does not exist. That is not a matter of judgement. There is no interface through which such a thing could be done.
What about the advertising product. A real thing. A separate thing.
OpenAI's documentation, checked on 30 July 2026, describes an OAI-AdsBot used to validate pages submitted as ads within ChatGPT. Paid placement is therefore a distinct route from being named as a source in an ordinary answer. Conflating the two is a way for a supplier to imply influence over organic answers that nobody has.
Why the absence of a dashboard matters so much. It removes verification.
In ordinary search a supplier's claims can be checked against a console the client also sees. Here there is nothing equivalent, so a client has no independent way to test what they are told. That is precisely the condition in which invented reporting flourishes, which is why our measurement guide exists.
What Appears To Influence It
Everything in this block is observation. OpenAI has not published how candidate sources are chosen or weighted when an answer is composed, so nobody outside the company can describe that. What follows is what practitioners see, offered as pattern.
Being described clearly on your own site. The most consistent pattern of the four.
Businesses whose own pages state plainly what they do, where they do it and for whom appear to be drawn on more readily than businesses whose sites communicate that through implication and imagery.
Being listed consistently across the web. The same details in every place they appear.
Where a name, address, description or service list contradicts itself between a website, a business profile and third party listings, an assembling system has conflicting material about who you are. Consistency appears to help. It also always helped ordinary local search, which is why our guide to consistency predates all of this.
Being mentioned by independent sources. Corroboration rather than self description.
A business described the same way by several unrelated sources appears more often than one described only by itself. The reasoning offered is that a lone claim cannot be verified. We find it plausible. Nobody has published it.
Being reachable by the crawler. The one item that is documented rather than observed.
Block four covers it separately, because it is the only thing on this page a business can verify for itself.
What we will not tell you. How these weigh against each other.
Any list assigning percentages to these has invented them.
Crawler Access Is A Real Lever
This is the one part of the subject that is neither inference nor opinion. It is published by OpenAI, it is verifiable on your own server. It is the only item here that can be got badly wrong by accident.
The documented position. Unusually explicit for this field.
OpenAI's crawler documentation, checked on 30 July 2026, states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links. It also recommends allowing that crawler, plus allowing requests from OpenAI's published address ranges, to help a site appear in search results.
Why this catches businesses out. Nobody blocked it deliberately.
Sites are excluded by broad security rules, by blanket restrictions inherited from an old configuration or by protective settings that treat an unfamiliar crawler as suspicious traffic. The site owner intended none of it and cannot see the consequence, since nothing on their own site looks wrong.
The controls are separate from each other. Which is the part most often misunderstood.
OpenAI's documentation, checked on 30 July 2026, states the settings are independent, so a business can permit the search crawler while declining the training crawler. Advice that treats blocking as a single yes or no decision misses that entirely.
There is a delay. Worth expecting.
The same documentation, checked on 30 July 2026, notes that it can take around twenty four hours after a change to a site's crawler instructions for OpenAI's systems to adjust for search results.
What we do about it. Check it, then keep checking.
We verify access as documented rather than assumed, confirm the traffic in server logs, then re-check after any hosting or security change. This is work rather than advice, which is why it is described here as work.
Why Answers Vary
Two people asking the same question about the same trade in the same town can be given different businesses. Nothing has gone wrong. No supplier can tune this away either.
Variability by design. The same question does not guarantee the same answer.
Generated replies involve an element of variation deliberately. Ask twice and the wording differs, sometimes with different businesses named, with nothing having changed anywhere.
Personalisation. Who is asking affects what comes back.
Location and the surrounding conversation both shape a reply. Two people are rarely running the same query even when they type the same words.
Memory of earlier exchanges. A stored context you cannot see.
Where a product retains earlier conversations, an answer can be shaped by things said weeks ago. That makes one person's result a poor guide to anybody else's.
Model updates. The system underneath changes.
Updates arrive on schedules nobody outside publishes in advance. A change on their side can move which businesses appear more than anything you did.
What this means for tracking a position. There is nothing stable to track.
A recorded position here describes one instance of a variable output. Repeat the check and the number moves without anything having improved. Our guide to how these systems assemble an answer covers why sampling over time is the only defensible approach.
How To Get Recommended
There is no mechanism for being recommended. There are conditions that appear to make it more likely. The distinction is not pedantry. A mechanism can be completed. Conditions can only be improved.
Condition one. You are reachable. The documented one.
The search crawler is permitted and confirmed arriving, per block four. Nothing else on this list matters if this fails.
Condition two. You have said what you do, plainly. In text rather than by implication.
What the business does, where it operates, what it does not do, what it costs where that can be stated. Written so a passage answers a question completely without depending on the paragraphs around it.
Condition three. Your details agree with themselves everywhere. Across the site, the profile and third party listings.
Condition four. Other people mention you. Independent references rather than self description, earned rather than bought, as our backlinks guides have always held.
Condition five. You are visible in ordinary search. Because one documented route to answers is a search index.
What no condition can deliver. Certainty.
A business meeting all five may still not be named for a given question. One meeting three may be. Anybody presenting this list as a method has changed conditions into a promise, which is the specific overclaim this cluster exists to avoid.
What We Would Actually Do
Deliberately unexciting, because the documented surface is small and we are not going to pretend otherwise.
Check access first. Against the vendor documentation rather than from memory.
Confirm the search crawler is permitted, confirm requests are arriving in the server logs, then re-check whenever hosting or security configuration changes. This is the only item with a documented consequence attached.
Fix what the business says about itself. Usually the largest piece of work.
Most sites we take on describe their services in language that assumes the reader already knows the trade. Restructuring so that individual passages answer real questions is the bulk of the job. It improves the site for people at the same time.
Make the record agree. Website, profile and listings saying the same thing.
Sample, then record. Rather than reporting a position.
A consistent set of questions asked on a schedule, with what appears recorded each time. That produces a trend over months. It does not produce a score. We will not invent one.
What we will not do. Three things.
Guarantee that ChatGPT names your business, since nobody controls that. Claim to know how sources are weighted, since OpenAI has not published it. Or charge you separately for work already covered by your ordinary search optimisation, since most of the above is exactly that work. Our approach page states those commitments in full.
One documented lever.
We start there.
OpenAI publishes that sites opted out of its search crawler will not be shown in ChatGPT search answers. Most businesses that fail this never blocked anything deliberately. It is checkable. Almost nobody checks.
What is included every month:
£350 per month, one target area. No setup fee, nothing billed separately.
Twenty guides.
One subject.
The mechanism, all four platforms, what you can influence, whether it is worth it yet, what it costs, how to measure it and how to judge a supplier selling it.
