How to improve your visibility in AI answers
There is a version of this article on almost every marketing blog, and it lists “the ranking factors for AI search.” Ignore every one of them. No AI engine publishes how it decides which brands to name. Not OpenAI, not Google, not Perplexity, not Anthropic. There is no documentation, no factor list, no confirmed weighting. Anyone printing one is guessing in a confident font.
What is left after you throw the guesses away is still worth doing, and it is more concrete than the guesses were. Some of it is mechanically certain: an assistant that cannot open your page cannot quote your page. Some of it is ordinary business hygiene that removes ambiguity about who you are. Some of it is inference from what you can watch these tools do — labelled below as inference, because that is what it is.
This guide is ordered by how sure anyone can be, strongest first. Do the first section even if you do nothing else on this page. Everything it asks for is free, and most of it is on our free tools shelf.
1. Be reachable — the only part that is mechanically certain
When an assistant answers a question by fetching pages, it sends an automated visitor to your website, reads what comes back, and may quote it. If that visitor is turned away at the door, nothing downstream matters. Your writing, your schema, your reviews — none of it is in the room.
This is not a theory about preferences. It is the plumbing.
- Check who you are letting in. Type your own address followed by
/robots.txtinto a browser —yoursite.com/robots.txt. That file is a plain-text note to automated visitors saying which parts of the site they may read. You are looking for lines that sayDisallowunderneath names likeGPTBot,OAI-SearchBot,PerplexityBot,ClaudeBotorGoogle-Extended. If you see those, the assistants those names belong to have been told to stay out. Our free AI crawler checker reads that file for you and lists what is allowed and what is blocked. - Know that this gets blocked by accident constantly. A website plugin with a “block AI bots” toggle switched on by a previous developer. A security setting at your CDN — the service that sits in front of your site to speed it up and filter traffic — set to challenge anything that is not a human. A staging site copied to live with its “hide from search engines” box still ticked. None of these announce themselves.
- Then check that your words arrive with the page. Some websites deliver an almost-empty page and then build the visible text in the visitor’s browser using JavaScript — code that runs after the page arrives. Whether a given assistant waits for that to finish is not published by any of them. What you can do is remove the question: if the text is already in the page as it is delivered, it does not matter what the fetcher does or does not wait for.
Who does what: you can do the robots.txt check yourself in thirty seconds, and reading it costs nothing. Changing it, undoing a CDN block, or getting text to arrive with the page are jobs for whoever built or maintains your site. Send them the checker result — it names the specific rule and file.
one honest note about direction
Unblocking a crawler does not cause anything to be quoted. Blocking one does prevent it. The relationship only runs one way — which is exactly why this is the first job and not the last.
2. Be unambiguous about who you are
An assistant answering “who does boiler repairs in Sheffield” has to decide whether the Miller Heating on your website is the same Miller Heating in the directory listing, the review site and the trade body register — or three different companies. When those sources disagree, nothing reading them — a person or a model — can tell which record is you without guessing. What any engine does after that guess is not published by anyone. What you can remove is the guess.
You control the agreement.
- One name, spelled one way. Not “Miller Heating,” “Miller Heating Ltd” and “MillerHeating” depending on where you filled in the form.
- One address and one phone number, identical across your website, your Google Business Profile, your social profiles and any industry directory you are listed in.
- One clear “who we are” page that states what you do, where you do it, and since when — in sentences, not adjectives.
- A machine-readable version of the same facts. This is a small block of code in your page called Organization schema — a standard format that states your name, address and official profile links so nothing has to be inferred from the layout. Our organization schema generator writes the block; you paste it in or hand it over.
Who does what: the name, address, phone and profile clean-up is yours, and it needs no technical help at all. The schema block is generated free above and then needs someone with access to your site to add it, which is a five-minute job for whoever built it.
stated honestly
No engine has confirmed that Organization schema causes it to name a business. What is true is narrower and still worth acting on: consistent, machine-readable identity removes an ambiguity that would otherwise have to be resolved by guessing, and inconsistency gives every reader, human or machine, a reason to pick the company it is more certain about.
3. Answer the questions buyers actually ask, in a form that can be quoted
Most business websites answer the questions the business wants to talk about. Assistants are asked the questions buyers actually have — about price, timing, suitability, what goes wrong, and what happens next.
Start offline. Write down the twenty questions you genuinely get asked on the phone, in your inbox, and in the first ten minutes of every consultation. Those are your prompts. Then answer them on your site, on the page where each one naturally belongs, with these habits:
- Make the first sentence stand alone. An answer beginning “As we mentioned above…” cannot be lifted out and reused. One that begins “Most repairs are finished the same day” can.
- Use specifics. Numbers, timeframes, materials, service areas, what is and is not included. “We pride ourselves on fast turnaround” says nothing. “Same-day for most repairs; two to three days when a part has to be ordered” is a fact a reader can use.
- Answer the awkward ones. What it costs, who you are not right for, what the common failure is. Competitors leave these out, which is precisely why they are yours to own.
- Publish them as visible questions and answers, not buried in prose. If you also want the machine-readable version, our FAQ schema generator builds it from the same questions you show on the page.
Who does what: all of this is writing, and you are the only person who knows the answers. Nobody needs to touch the code.
stated honestly
None of this is a promise that an answer will be quoted. There is no published rule saying a self-contained sentence is preferred. The claim here is smaller, and it is about your text rather than about the engine: an answer that can only be understood by reading the three paragraphs above it is harder for anything to reuse intact — a person skimming, a colleague pasting it into an email, or a model summarising.
4. Be present where the engines already look — labelled as inference
weakest evidence on this page
Which sources an engine trusts is not published by anyone. What follows is inference from what you can watch happen, not a mechanism. Every competing article states this section as fact; it is not one.
Here is the observable part, and you can do it this afternoon. Ask ChatGPT, Perplexity and Gemini the questions your buyers ask. When an answer shows its sources, read the list. Those are pages that were actually fetched for that question, on that run. You are not learning a ranking rule; you are learning which corners of the internet came up when your category was discussed.
Usually it is some mix of industry directories, trade associations, local and trade press, comparison and roundup pages, and forums where your category gets argued about. If your business is absent from all of them, being present is a reasonable thing to work on: a directory listing, a trade body membership, an accurate profile on the review site your industry actually uses, a genuine contribution to the place your customers already read.
Two cautions, both from reading the evidence honestly. First, source lists vary between engines, between questions, and between two runs of the same question — a source that appeared once is not a rule. Second, anyone selling you “AI placement,” guaranteed mentions, or a link package that promises AI visibility is selling something they do not control. There is no placement to buy inside an organic answer.
5. What you cannot control, said plainly
Half the value of an honest guide is the list it refuses to sell you.
what a model already learned
Part of what these systems say comes from training, not from fetching anything. If your company launched after a model’s training data was collected, that model does not know you exist, and nothing you publish this week changes what it already learned. Being fetched at answer time is the one route that does not wait for a retraining cycle you cannot see or schedule — which is section 1 again.
surfaces with no API
Some AI answer surfaces have no programmatic route at all, so they cannot be tracked — only spot-checked by a human. Anything claiming otherwise is describing a scrape, or a person looking occasionally. The detail is in our guide to what can actually be measured.
run-to-run variation
Ask the same question five times and you can get five different answers. This is how these systems generate text, not a fault, and it means a single check is closer to a coin flip than to a measurement.
your own logged-in view
The answer you see in your own account is shaped by your history and your settings. It is not what a stranger sees, and it is not evidence in either direction.
Nobody outside these companies knows the weighting, and the companies do not publish it. A guide that hands you a factor list is inventing the part it does not know.
Who does what
Four of the five jobs on this page need nobody but you. Only the first two touch the site itself.
| The job | You can do this yourself | Hand to whoever built your site |
|---|---|---|
| 1. Reachability | Open yoursite.com/robots.txt and run the free crawler check | Editing that file, lifting a CDN or plugin block, getting your text into the page as delivered |
| 2. Identity | Fix the name, address and phone everywhere; claim your official profiles | Pasting in the generated Organization schema block |
| 3. Buyer questions | Write the answers — you are the only person who knows them | Nothing, unless adding a page to your site is awkward |
| 4. Presence | Claim the directory, trade-body and review listings that already exist | Nothing |
| 5. Measuring | Ask your question set on a schedule and record the results | Nothing |
How to tell whether any of it worked
Fix things in the order above, then measure like this: take the same fixed set of buyer questions, ask them repeatedly on a schedule, and record how often you are named across the runs. Report the spread — “named in three of five runs” — never a single result, and never a screenshot of one good answer.
The one exception is section 1, which has a before-and-after you can actually see: blocked yesterday, not blocked today, checkable in a file. Everything below it is a rate that only means something across weeks, and only if the question set stays the same between the dates you are comparing. Change the questions and the comparison is void.
What the numbers can and cannot be made to mean is its own subject — the metrics guide covers the difference between being named from memory and having your page actually fetched, which decides which half of this page you should be working on. The step-by-step manual routine has its own guide rather than a paragraph here.
FAQs
How long does it take to improve visibility in AI answers?
Nobody can tell you honestly, because no engine publishes how often it refreshes what it knows or retrains. What is quick to verify is the technical half: whether an assistant can reach your pages is checkable today, and a block removed today is removed today. Whether that changes how often you are named is only visible as a rate across repeated checks over weeks.
Can I pay to appear in ChatGPT or Perplexity answers?
Not in the answer itself. There is no placement to buy in an organic AI answer, and any agency offering guaranteed mentions is selling an outcome it does not control. Paid advertising, where an engine runs it, is a separate and labelled surface, not the recommendation in the answer text.
Does adding an llms.txt file improve AI visibility?
No major AI vendor has confirmed that it reads llms.txt. It is a proposed convention, not a supported standard, and it costs nothing to add — so treat it as a cheap checkbox, not as a lever. Any guide telling you it makes AI read your site is describing a wish.
Will more schema markup get my brand mentioned?
No engine has published that it will, so nobody can promise it. What schema does is state your facts in a format that does not have to be interpreted from your page layout, which removes a source of error. That is worth doing on its own terms, and it is not the same as a ranking mechanism.
Do I need to publish content constantly to stay visible?
There is no published rule linking publishing frequency to being named, and volume for its own sake mostly produces pages nobody asked for. The version of this that holds up is coverage: every real buyer question answered somewhere on your site, specifically, and kept accurate as your prices and services change.
My competitor gets named and I don’t. What do I check first?
Whether an assistant can reach your site at all — the robots.txt file and any CDN or plugin blocking automated visitors. That is the one failure with a definite answer. If you are reachable and still absent, the next question is whether you are being named without your site ever being fetched, which is a different problem with a different fix.
Which buyer questions name your competitors and not you?
The same questions, asked once each of ChatGPT and Gemini and reported as one run — free.