How do you improve brand visibility in AI answers?

This is the whole-picture version: what to do, in what order, and what to expect from each. The engine-specific detail lives in the guides for ChatGPT, Perplexity and AI Overviews.

First, find out which half is broken

There are exactly two reasons an assistant does not name you, and they have opposite fixes. Doing the wrong one is the most common waste of money in this field.

They cannot read you. Content that only exists after JavaScript runs, a blocked crawler, no structured data, an entity the model cannot pin down. This is cheap to check and comparatively fast to fix — a free scan will tell you in a minute.

They can read you fine and still do not name you. Your site is immaculate and you are absent from the answers anyway. This is the more common case, and no amount of further website work will move it.

The reason to establish which one you are before acting: a perfect site that no assistant mentions is not a website problem, and the entire "optimise your pages" industry will happily sell you the wrong remedy.

If they cannot read you — the fast work

Weeks, not months, and mostly one-off:

• Serve your content in the HTML. If the words only appear after client-side rendering, assume a crawler saw an empty page.
• Let the crawlers in. Check robots.txt for AI user-agents specifically, and check your CDN's bot rules, which are a more common cause of a blocked crawl than robots.txt is. A site that cannot be fetched cannot be read, and we cannot even record a reading for it.
• Be an unambiguous entity. Consistent name, address and description everywhere you appear, plus Organization schema. Models resolve entities; give them one thing to resolve to.
• Answer questions in prose. Extractable passages — a direct answer near the top, plain declarative sentences — are what gets lifted into a summary.

If they can read you — the slow work that actually matters

Assistants decide who belongs in an answer largely from other people's pages. When we look at the sources cited alongside a category recommendation, they are mostly not the recommended brand's own website — they are roundups, comparisons, retailer lists, forums and trade press.

We learned this in an embarrassing way. An early version of our remediation advice read the cited sources for a category and told brands to get themselves mentioned on those domains — some of which turned out to be competitors' sites. The bug was ours, but the underlying signal was real and stands: the pages shaping who gets recommended are overwhelmingly third-party, and they include sites nobody would have put on a competitor list.

So the work is:

1. Get into the roundups. "Best X in the UK" listicles, category comparison pages, buyer's guides. These are what a live web search retrieves and what a training corpus is thick with. One inclusion in a well-cited roundup does more than a quarter of blog posts on your own domain.
2. Get covered by trade and local press. Independent description of what you do, on domains models trust.
3. Be present where your category is discussed. Forums, review sites, community threads — as a genuinely useful participant. Astroturfing is both detectable and against most platforms' rules, and it is not what this means.
4. Fix what is said about you where you already appear. An out-of-date listing describing a business you no longer are will be repeated back confidently for years.

Expect this to be slow. Retrieval-led answers can pick up new coverage within weeks; anything drawn from training memory waits for a model release.

What we cannot show works

We measure this for a living and we still cannot demonstrate a clean effect from several popular tactics. Treat them as unproven rather than useless:

• llms.txt — no assistant has committed to reading it, and we have not seen a reading move because of one. Cheap to add, so add it if you like, but do not count it as work done.
• Schema as a ranking factor. Structured data demonstrably helps a machine understand you. That it increases how often you are recommended is a separate claim, and one we cannot yet evidence.
• Volume of your own content. Publishing more pages about yourself is the intervention most often sold and the one with the weakest link to being named by an assistant.

Knowing whether it worked

Re-measure the same way. Same questions, same assistants, same counting rules — change any of them and you have measured your method rather than your brand. And hold the result against its confidence interval before calling it progress: a score that moves from 8 to 9 has probably not moved at all.

How to run a repeatable audit sets out the procedure, and how a score is calculated explains the range you should be comparing against.