How to improve brand visibility in AI search

The short version

Improving AI visibility is mostly not a content exercise on your own website. It is: measure where you are absent, find what the assistants read instead, and get your business represented accurately in those sources.

The order matters more than the effort. Without a baseline you cannot tell a change from noise — answers drift by several points on their own — and without knowing which sources decided the answer, you will spend the budget on the page that felt most important rather than the one that mattered.

Step 1. Measure before you change anything

One check is an anecdote. Assistants do not answer identically twice, so a measurement means the same buying questions asked several times per assistant, reported as a share rather than a yes or no.

Without that baseline, the number you get in six weeks answers no question, because compared to what? has no answer. This is the step people skip because it produces no visible work, and it is the one that makes every later step interpretable.

Measure per assistant, never averaged. Four of them asked the same questions in one market named 63 businesses between them, and only four were named by all four; 46 appeared in a single assistant only. A blended number would have described none of them.

Record three things: whether you were named, who was named instead, and what was cited. The second and third are more actionable than the first.

Step 2. Find the question types where you are absent

Visibility is rarely uniform. A business is often named in general recommendation questions and invisible in price or urgency questions — or the reverse.

Group the questions by what the buyer wants:

• general recommendation — who should I use for X in this city
• specific service — who does Y
• price — how much does Y cost, who is affordable
• urgency — who can see me today
• trust — who is reputable, who do people recommend
• comparison — X or Y, which suits me

The group you are absent from is the work. Fixing a group you already win changes nothing, and it is the most common way a programme produces effort without movement.

There is also a commercial hierarchy here. Price and urgency questions sit closer to a decision than general recommendation questions. Being absent from “who is open now” costs more than being absent from “best in the city”, even though the second sounds more prestigious.

Step 3. Look at what was read instead

This is the step with the leverage, and the one that gets skipped because the output is a list of pages rather than a list of opinions.

In one market we measured, a single assistant answer rested on eleven pages on average — the business’s own pages alongside directories, review sites, local roundups and trade listings.

Read that list as a work order in two parts. The pages of yours that were cited, or should have been, are the ones to rewrite first: they are the only description of you that you control, and everything else copies from them. The rest of the list is who has to be made to agree with you.

How to build the list. Where citations are shown — Perplexity always, AI Overviews usually — write down every cited domain for the questions you care about. Where they are not shown, the pattern across answers still points at the source types that carry weight in your category: the same three or four names will keep appearing in the text.

What you will find. Usually something duller than expected. Four directories, a review platform, one or two regional roundups, and occasionally a trade association page nobody has updated since 2021.

Step 4. Make the facts agree everywhere

Assistants cross-check sources. Services, hours, address, specialities and pricing that differ between your site and three directories give a model less reason to name you than a business whose description is identical everywhere.

This is unglamorous, it does not photograph well in a report, and it is most of the work.

A practical checklist for each source on your list:

1. Are you there at all? The most common finding is a missing listing rather than a wrong one.
2. Is the business name spelled exactly as on your own site? Including the legal suffix or its absence — pick one and use it everywhere.
3. Are the services listed, in the words buyers use? Not your internal service codes.
4. Are prices there, if you publish them? A source with prices is more useful to an engine answering a price question, which is precisely where most businesses are absent.
5. Are the hours right, including the exceptions? Urgency questions depend on this.
6. Is the description the same description? Not a rewritten one — the same claims, the same specialities.

Step 5. Make your own pages answerable

Where your pages are read, what matters is whether they answer the buying question in the first two sentences, in the words a buyer uses.

Use the question as the heading. Not Our Services but Where do I go for an emergency appointment on a Sunday?

Answer it immediately. The first two sentences under a heading should answer that heading completely. Everything after is supporting detail for the human who stayed.

Make each section stand alone. Define terms in place. Repeat the noun rather than writing “it” across a paragraph break. An extracted passage has no surrounding context to resolve.

Use structure a machine can parse. Lists, tables as HTML rather than images, consistent heading levels.

Step 6. Publish something checkable

Original numbers, stated method, real examples, dates. Assistants prefer a source that says something specific to one that says something impressive, and so do the people reading the answer.

This is the slowest lever and the most durable. A page with a number in it that nobody else has gets cited for years; a page that says “leading provider” gets cited never.

If you have data nobody else has — even small data, even about your own customers — that is the highest-value page you can publish.

Step 7. Measure again, and be honest about the result

Same questions, same way, same assistants.

Movement under about five points is drift, not progress. Answers vary that much on their own. Calling a four-point rise a win is how agencies end up reporting weather as work, and it is the single fastest way to lose a client’s trust when they eventually check for themselves.

Look at where the movement landed. If the share rose in the question group the work targeted and stayed flat elsewhere, that is weak evidence the work did something. If it rose evenly everywhere, that is more likely drift or a model update.

If nothing moved, say so and go back to step 3. Usually the source list was incomplete, or the listings were corrected but not yet re-crawled.

How long does each step take?

The honest total is four to eight weeks before a second measurement is worth running, and the gating factor is almost never the measurement — it is getting a straight answer from inside the business about what it actually offers and for how much.

A worked example

From a measurement on one local market, business anonymised.

Baseline. Twelve buying questions, five repeats, four assistants. Named in 38% of general recommendation answers and 0% of price answers.

The gap. Price questions, entirely. General recommendation was healthy.

What the citations showed. For price questions, the same four sources came back repeatedly: two directories, a review platform, a regional roundup. The business was on one of the four, with an out-of-date service list and no prices.

What was done. Nothing to the homepage. Listings corrected and completed on all four, using the same service names and price ranges as the business’s own pages. One page rewritten so the price question is answered in the first two sentences rather than in a table three screens down.

What moved. General recommendation shifted four points — reported as no change. Price answers went from zero to being named in roughly a third of them.

What was not claimed. That the listings caused it. What was reported: what changed, when, and what moved after, with the note that movement was concentrated in the targeted question group rather than spread evenly.

Mistakes that waste the most money

Starting with the website. It feels like the obvious lever and it is usually not the one that decides the answer.

Measuring once and drawing conclusions. Or worse, measuring once before and never again.

Reporting an average across assistants. Produces a number nobody can act on.

Chasing the assistant that looks worst. Shares move independently; a programme tuned to last month’s weakest engine will reverse itself next month.

Treating traffic as the measure. An assistant may name you without linking; the visit arrives days later as direct traffic with no explanation.

Buying a guarantee. The lists are not published, not stable and not for sale. Anyone promising placement is guessing.

What to expect, honestly

You can expect to know where you stand, which question types you lose, which sources decide them, and what moved after the work. That is a real and unusually checkable set of things to know.

You cannot expect a promise about placement, a fast result, or certainty about cause. We can put dates next to each other; claiming causation from that would be dishonest, and the category has enough of that already.

A checklist for the first two weeks

Print this, work down it, and stop when you run out of sources rather than when you run out of enthusiasm.

Week one

1. Write ten buying questions the way your customers would ask them.
2. Ask each one five times, in each assistant you care about. Record the answers verbatim.
3. Count the share that named you, per assistant, per question type.
4. List every business named instead of you.
5. List every source cited, by domain.

Week two

6. Check whether you appear on each cited source. Note which are missing you entirely — those are first.
7. Compare what each source says about your services against your own site. Write down every difference.
8. Correct the listings, starting with the sources cited most often.
9. Rewrite one page so that the buying question you lose most is answered in its first two sentences.
10. Diarise the second measurement for four weeks out, not two. Listings take time to be re-read.

What to do when nothing moves

It happens, and the response matters more than the result.

Check the listings were actually re-read. Corrected pages take weeks to be crawled again. Measuring after two weeks and concluding failure is the most common false negative here.

Check the source list was complete. If citations were only visible on one surface, the list you worked from may have missed what the others read.

Check the question set still matches how people buy. Markets move, and a set written six months ago can be measuring questions nobody asks.

Then say so. A report that says “nothing moved, here is what we checked, here is what we think is blocking it” is worth more than one that finds four points of drift to celebrate.

Where to start

• The free AI visibility check covers step 1 for one business at no charge, and emails you what the assistants said.
• The method explains how the question set is written and why every figure is per assistant.
• The sample report is open in full, with no form.
• Background: what AI visibility is, and the full AI search optimization guide.