How to Improve Your Visibility in ChatGPT

What does visibility in ChatGPT actually mean?

It means how often your business is named when people ask the assistant questions that your business could answer. It is a frequency measured across a controlled set of prompts, not a position, and there is no ranking to occupy.

The phrase ranking in ChatGPT gets used constantly and there is nothing for it to refer to. No index is ordered and no position exists to be won. What exists is a percentage: of the relevant questions you decided to measure, the share in which you were named, and whether that share is going up.

Three outcomes hide inside the word visibility, and they need separating before any of the rest of this article is useful. A recommendation is the assistant putting your business forward as the thing to choose. A mention is your name appearing in the answer text, which may be neutral, comparative or unflattering. A citation is a source link attached to the answer, which may point at your website or at somebody else's page that happens to describe you.

These move independently, and the gap between them is large enough to matter commercially. Being cited without being mentioned is common, as is the reverse. Most of the practical work below improves one of the three more than the others, which is the reason to know which one you are short of before starting. And throughout, the shape of the problem is worth accepting: you are trying to make yourself a more probable answer, not a guaranteed one.

Why the answer can change

Ask the same question twice and you can get two different sets of businesses. This is normal and it is not a fault in your setup.

Seven things move underneath a given answer, and it is worth knowing which is which, because they have different implications for how you test:

• Prompt wording. The largest single source of variation, and the one most people fail to control. Asking for the best plumber and asking who to call about a leak can return different businesses, or none at all.
• Conversation context. Everything already said in the thread informs the answer. A question asked cold and the same question asked after three exchanges are not the same input.
• Location. Whatever the system has, whether stated explicitly or inferred from the session, and it changes local answers completely.
• Retrieval mode. Whether live material was fetched or the answer came from what the model already held. This is not exposed to the user.
• Model and version. Products are updated without notice, so a comparison across several weeks may be comparing two systems rather than two states of your business.
• Reasoning mode. Where a product offers a faster and a more deliberate mode, they behave measurably differently, including in how much they cite.
• Probabilistic generation. Even with everything above held constant, responses are generated rather than retrieved from a table, so repetition alone produces variation.

The practical consequence is that a single screenshot proves nothing, in either direction. If someone shows you one answer as evidence of a problem or a success, they are showing you a coin flip.

More striking than either of those numbers is how little the two modes agreed about where to look. Across the same prompts, only a quarter of the cited domains were common to both.

That result is worth sitting with, because it undermines a common assumption about how this work is done. If two modes of one model on one set of questions agree on a quarter of their sources, then a source list assembled by watching a handful of answers is not a plan. It is a sample, and a small one.

The practical response is not despair but sample size. Variability of this kind averages out across a large enough prompt set run repeatedly, which is exactly why the measurement discipline at the end of this article is not optional bureaucracy.

Website crawlability

Start here, because everything else is wasted if this is broken. If the relevant crawlers cannot reach your pages, no amount of content or reputation work will help.

The reason this matters more than it did is that a retrieval step is not a person. A human who hits a page that renders badly waits a moment, or reloads, or works out that the content is behind a tab. A fetch either returns usable text or it does not, and nothing downstream recovers from the second case. Accessibility to a retrieval system is binary in a way accessibility to a reader is not.

The audit is short and mostly mechanical:

• robots.txt. Which user agents are allowed, and whether anything is being blocked by an old rule nobody remembers adding.
• Crawler policies. Any blocking at the CDN, firewall or bot-management layer, which is where most unintended blocks actually live rather than in robots.txt.
• Status codes. Important pages returning 200 rather than a redirect chain, a soft 404 or an intermittent 5xx under load.
• Canonical URLs. One canonical version per page, agreeing with internal links, the sitemap and whatever the CMS emits.
• noindex and nofollow. Directives left behind from a staging environment or a template, which is more common than it sounds.
• Server-rendered content. The substance of the page present in the initial HTML, not assembled after scripts run.
• JavaScript dependency. Whether the answer to the page's question survives with scripting unavailable. Tabs, accordions and lazy-loaded sections are the usual offenders.
• Internal links. Every page you care about reachable from somewhere, in ordinary anchor tags rather than script-driven navigation.
• Orphaned pages. Pages with no internal links at all, which happens routinely with landing pages built for campaigns.
• XML sitemap. Present, current, and not listing URLs that redirect or no longer exist.
• Interaction requirements. Whether the content is genuinely available without accepting a banner, dismissing a modal or clicking to expand.

That last point deserves emphasis because it catches otherwise well built sites. A pricing table inside a collapsed accordion, or a service description that appears only after a tab is clicked, may be perfectly present to a reader and effectively absent to anything fetching the page.

On the question of which crawlers to allow, OpenAI publishes its user agents and how site owners can control them, and it is worth reading that documentation directly rather than a summary of it, because the agents and their purposes have changed more than once. Verify any crawler-specific claim against the current official documentation before acting on it, including the ones in older articles.

It is a decision worth making deliberately rather than by default. Some publishers do block these crawlers, and that is a legitimate choice with a predictable cost: you are trading the possibility of being cited for control over how your content is used. What is not defensible is discovering two years later that a blanket block was inherited from a template.

Search availability

Because assistants can retrieve live search results, being findable in conventional search is one of the more dependable routes into being mentioned.

This is the part that most reliably rewards existing SEO work. If your pages are indexed, rank for the non-branded queries your customers use, and answer those queries clearly, you are already in the pool of material a retrieval step can draw on.

The important nuance is that a page does not need to rank for the exact wording of somebody's prompt in order to be retrievable for it. A conversational question is a poor search query, so systems tend to reformulate: one prompt can become several retrieval queries, and the material that comes back may have been found through a synonym, an adjacent question, a narrower sub-question, a supporting fact or a broader concept than the one that was typed.

This is sometimes called query fan-out, and the honest position is that its mechanics are not documented for any major assistant. What is observable is the consequence, which is well evidenced: cited pages frequently do not rank for the literal prompt. We look at the data on that in the AEO and SEO comparison.

Practically, this changes what a content gap looks like. The question is not only whether you rank for a phrase but whether anything you publish would be retrieved by any of the plausible reformulations of a customer's question. A page that covers a topic thoroughly and specifically has many routes in. A page built around one exact-match phrase has one.

It also means that if you have been neglecting conventional search on the assumption that it no longer matters, you have probably made your AI visibility worse rather than better.

Entity clarity

A model has to work out what your business is before it can decide whether to recommend it. The easier that is, the better.

Entity clarity means the basics agree wherever they appear: the legal and trading name, the address, the service area, the category of business, the services offered and how to make contact. If your website says one thing, your map profile says another and an old directory entry says a third, you are asking a system to resolve a conflict, and the cheapest resolution is often to recommend somebody less ambiguous.

The pattern is easier to see with a worked example. Suppose an illustrative dental practice, which we will call Harbour Dental, presents itself as follows. Its website is headed Harbour Dental. Its business profile is registered as Harbour Dental Clinic Ltd, because that is the legal entity. A directory entry created during a rebrand five years ago still says Harbour Cosmetic Dentistry, and lists the previous premises two streets away, which is also still on two aggregator sites that syndicated it at the time.

None of that is a mistake anybody made. It is the ordinary residue of a business changing over five years. But consider what it looks like from outside: three names, two addresses, and a category (cosmetic dentistry) that no longer describes what the practice mostly does. A system trying to answer a question about emergency dentists in that area now has to decide whether these are one business or three, which address is current, and whether the entity it can most confidently identify actually provides the service being asked about.

The fix is not clever. It is choosing one canonical name, deciding what the old names redirect or defer to, correcting the address everywhere it appears including the aggregators, updating the primary category, and then checking the same list again in six months because syndication reintroduces old data. This is dull work with an unusually good return, and it is the area where we most often find something genuinely broken.

Content usefulness

Write the page that answers the question, then answer it immediately. That is the principle, and stated on its own it is too abstract to act on, so the rest of this section is what it actually requires.

1. Name the subject explicitly

Do not make a reader or a retrieval system infer what it, this, we or our solution refers to. Where a sentence could plausibly be about the company, the service, the industry or the previous paragraph's example, name which one it is.

This is not an instruction to repeat the company name in every sentence, which reads badly and helps nothing. It is an instruction to check the sentences that carry the important facts, because those are the ones that will be read out of context.

2. Answer before elaborating

If the heading asks how much Invisalign costs in Austin, the first paragraph should give a range, or explain specifically why no range can be given and what the price depends on. It should not open by observing that every smile is unique and arrive at the number six paragraphs later, under a subheading.

The instinct behind the delay is usually commercial: the writer wants the reader invested before the price appears. It does not survive contact with how people read online, and it definitely does not survive a system looking for the passage that answers a question.

3. Include facts a general model cannot know

This is the highest-value paragraph in the section, because it is the part a language model cannot supply for itself. A competent model can already explain what a service is. It cannot know:

• your price, or an honest range with the variables named
• your service area, stated in place names rather than implied by an address
• your opening hours, and whether you handle urgent work outside them
• turnaround or waiting times, realistically
• what the process actually involves, step by step
• eligibility, prerequisites and who a service is not suitable for
• the qualifications and registrations of the people doing the work
• current availability, where that is a real constraint
• what is excluded, and what commonly surprises people
• what happens at a first appointment or during onboarding
• what a package or tier includes, specifically
• how your version differs from the two or three alternatives a buyer is weighing

A page containing six of those is doing something no amount of general explanatory content can substitute for. Most service pages contain none of them.

4. Make useful passages self-contained

A good paragraph should retain its meaning when extracted from the page. The test is mechanical: read it on its own and see whether it still says something checkable.

Again, this is not a licence to write robotically. It means that the paragraph carrying your pricing should not depend on a subheading two screens up to establish which service is being priced, and the one describing your service area should not rely on the reader having noticed the city in the page title.

5. Organise around real questions

The source material for this is already inside the business, and it is better than anything a keyword tool produces. Recorded sales calls, the questions support answers repeatedly, the objections that surface before a decision, the comparisons prospects raise unprompted, the phrasing that appears in reviews, and where relevant the way people discuss the category in communities and forums.

What that gives you, and keyword research does not, is the qualifiers. A tool returns implant cost. A sales call returns whether implant cost changes if you need a bone graft first, which is the actual question and the one nobody has written a page about.

The failure mode to avoid is the opposite: generating a hundred near-identical FAQ pages from keyword variations. That produces a site that is worse for readers, thin from every other perspective, and no more likely to answer anything.

6. Prefer specificity to volume

One page that answers an important commercial question completely is usually worth more than ten pages covering minor variations of it. This is the trade most content plans get backwards, because volume is easier to brief, easier to measure and easier to sell.

7. Use page format deliberately

Different kinds of page carry different retrieval value depending on the question. A pricing page, a genuine comparison, a service page and a clearly structured explanation are not interchangeable containers for the same words, and the published research suggests format matters independently of how well a page is written. The evidence, and its limits, are in what makes AI cite a page.

The practical question is whether the commercially important questions in your category have a page of the right shape at all. Many businesses have a blog and no page that states what anything costs.

8. Write for people, and make the answer recoverable

This is the synthesis, and it is less of a compromise than it sounds. The prose should read naturally, because humans are the ones who buy. The structure should make the meaning easy to recover, because that is what makes a passage usable by anything reading it, including a person skimming.

The two goals only conflict when somebody is writing for machines, at which point both fail.

Third-party authority

What other sites say about you carries weight that your own site cannot carry on its own. This is the uncomfortable part of the work, because it is the least controllable.

The realistic version is not a link building campaign. It is making sure you are present and accurately described in the places that already matter for your category: professional bodies, trade associations, legitimate industry directories, local press, supplier and partner sites, and any accreditation you actually hold.

Quality matters far more than volume here, and low-quality directory submissions are largely a waste of money. One accurate entry on a body that genuinely governs your profession is worth more than fifty listings nobody consults.

Citations

Where an assistant shows the sources behind an answer, those sources tell you what it currently trusts on that topic. This is the most useful free diagnostic available, and almost nobody reads it systematically.

The workflow is worth doing properly, because done casually it produces a list of domains and no decisions. For each important prompt:

1. Record whether you were named, and separately whether you were recommended or merely mentioned. These are different results.
2. Open every cited source. Not the domain, the actual page. The domain tells you very little about why it was used.
3. Categorise it. Your own site, a competitor's site, a review platform, a directory, a professional register, trade press, a comparison article, a forum thread, something else.
4. Record recurring domains across prompts. A domain that appears once is noise. One that appears across a third of your prompt set is part of your category's evidence layer.
5. Identify what each page actually supplies. Prices? A list of providers? A definition? Reviews? Credentials? This is the part that turns a list into a diagnosis.
6. Check whether you appear there, and accurately. Present, absent, present but wrong, or present but describing a version of the business from four years ago.
7. Note the missing source types. If every cited source in your category is a review platform and you have thirty reviews, that is the finding.

What usually emerges is a small set of sources doing most of the work in a category, and it is frequently not what the business expected. Sometimes it is a review platform, sometimes a licensing body, sometimes a single well-written comparison article on a site nobody in the industry had heard of. Once identified, a vague objective becomes a concrete list: be present, be accurate, and understand why those sources are being trusted.

One caution on scope. The set of sources differs by engine and, as the reasoning-mode research earlier in this article showed, can differ substantially within a single product. A source list built from one assistant is a starting point rather than a map.

The strategic reading is about where effort pays off. Where a large share of cited URLs are brand-controlled, improving your own pages is doing real work. Where most of them are third party, the pages that decide the answer are ones you do not own, and the useful effort goes into being accurately represented on them instead. Those are different budgets and different briefs.

Diagnosing an absence

When a business is not appearing, the first job is to establish which kind of absence it is, because they have almost nothing in common except the symptom. Work down these questions in order and most cases resolve within the first four.

1. Was anything business-specific returned at all? If the answer was general advice with no companies named, this is not a visibility problem. The prompt does not produce recommendations, and no amount of work on the business will change that. Fix the prompt set.
2. Are competitors appearing? If yes, the question produces recommendations and you are losing a comparison. If no competitors appear either, the category or the geography may simply be poorly documented, which is an opportunity rather than a deficit.
3. Is the site indexed and retrievable? Check the mechanical things before the interpretive ones. A blocked crawler or a JavaScript-only service page explains a total absence better than any theory about authority.
4. Are the competitors who appear better documented, or just bigger? Read their pages against yours on the specific question asked. Frequently they have stated a price, a service area or a qualification that you have not.
5. Does the business visibly provide the requested service? Not implicitly, in words, on a page. A practice that does emergency work but describes itself only as a general practice will lose emergency questions.
6. Is the geography explicit? Named towns, neighbourhoods and service areas in text, not an address in a footer and an assumption.
7. Are the dominant cited sources third-party platforms you are missing from? If the answer is being assembled from three directories and a review site, being absent from those is the whole explanation.
8. Is the reputation profile too thin or too stale to support a recommendation? A handful of reviews from three years ago gives a system very little to justify naming you with.
9. Is the entity inconsistent? Two trading names, a moved address, a superseded category. This is the most common single cause we find and the least likely to be suspected.
10. Is this simply variance? Ask again, several times, on several phrasings, before concluding anything. A business named in three runs out of ten is not absent.

Reputation, locality and structured data

These three do less on their own than the sections above and more in combination, which is why they belong together. None of them will make an absent business appear. All of them make a business that is already plausible easier to justify naming.

Reviews and public reputation

Reviews function as evidence a system can point to when justifying a recommendation, which makes them more important here than in conventional search. Volume, recency and rating all appear to matter, though we would be cautious about anyone stating the exact weighting, because we have not seen a credible study establishing it and we have not finished our own. What is clearly true is that a business with almost no reviews gives a system very little to work with.

The text is worth attention too. Reviews that mention the specific service, the location and the outcome give a model far more usable material than reviews that only say the experience was good. There is more on how reviews function in local answers specifically in how AI systems find and recommend local businesses.

Local signals

If you serve a specific area, local signals do a large amount of the work, because most commercially valuable questions have a location attached. The essentials are a complete and accurate map profile in the right primary category, an address and service area that match everywhere else, opening hours that are actually correct, and pages on your own site that name the areas you serve without descending into a list of a hundred near-identical location pages.

Local intent also narrows the field considerably, which is good news: it is far easier to become a consistent answer for one city and one service than for a whole country.

Structured information

Structured data does not make a business recommended. It makes a business easier to interpret without guessing, which is worth having. For most businesses the useful set is small: an Organization or LocalBusiness description, the services offered, and FAQ markup where you genuinely have questions and answers on the page. Keep it accurate and consistent with the visible content, because markup that contradicts the page is worse than no markup.

Treat it as removing ambiguity rather than sending a signal. That framing keeps expectations in the right place, and it is consistent with the citation research, which found no independent effect from markup once other factors were controlled for.

Measuring visibility

Build a fixed prompt set and re-run it on a schedule. Everything else is anecdote.

A workable set is somewhere between fifty and a few hundred questions, written the way customers actually ask rather than the way marketers write keywords. Getting the composition right matters more than getting the size right, and a set that covers eight kinds of question is more informative than one covering three hundred variations of one:

• Informational. What a service is, how something works, what a term means. These often produce no companies at all, which is itself worth knowing.
• Recommendation. Who should I use, who is best, who do you suggest. The prompts that produce shortlists, and the core of the set.
• Comparison. This option against that one, or one named provider against another.
• Transactional. Pricing, booking, availability, how to get started.
• Location-based. The same questions with a place attached, at several levels of granularity.
• Service-specific. The individual named services rather than the category.
• Reputation. Is this provider any good, what do people say, is it trustworthy.
• Problem-led. The symptom rather than the solution, which is how a lot of people actually ask. Somebody with a leak does not search for plumbers.

Branded prompts should be a small minority of the set, and they should be labelled as such when reporting. Asking an assistant about your own company and being pleased that it knows who you are is not a visibility measurement: you have supplied the answer inside the question. A prompt set dominated by branded questions produces a number that looks excellent and moves in response to nothing.

Run the set across the assistants your customers plausibly use, record whether you were named, and record who was named instead. Track the share of prompts naming you, your share against named competitors, and what gets cited. Then track whether enquiries are moving, because that is the only number that pays for the work.

Common mistakes

Nearly all of them are measurement errors rather than execution errors. The work itself is not especially hard to get right; drawing conclusions from it is where most of the money gets wasted.

• Judging by a single answer. One response tells you almost nothing.
• Asking a leading question. Typing your own company name and being pleased it appears is not a test.
• Abandoning SEO. In several products search visibility is the retrieval path.
• Publishing volume. Large amounts of thin content aimed at models is expensive and tends not to work.
• Ignoring the profile. An out-of-date listing quietly undermines everything else.
• Buying guarantees. Nobody can guarantee inclusion in a generated answer.

What to do first

In order, and without skipping the first one:

1. Establish a baseline. Write thirty to fifty real customer questions and run them. Record what happens.
2. Check crawlability and decide deliberately which crawlers you allow.
3. Fix entity consistency. Name, address, category and services, everywhere they appear.
4. Look at what gets cited for your questions, and make sure you are present and accurate there.
5. Fill the obvious content gaps, answering each question directly at the top of the page.
6. Work on reviews, steadily rather than in a burst.
7. Re-run the same prompt set a month later and compare like for like.

If you would rather not assemble that yourself, our free audit is essentially step one done for you.