How to Choose Your First AI Visibility Optimization and Prove It Worked

How to choose your first AI visibility optimization: read your baseline counts to pick one change, skip what Google says it ignores, log each test and judge it against a range.

Angelique Swain · · 7 min read

Cover reading How to Choose Your First AI Visibility Optimization, beside ranges of 15% to 26% and 29% to 42% for 40 and 70 named in 200 answers
On this page

You've been tasked with AI visibility optimization. Where do you start? Begin by counting how often AI answers name and link your brand for your buyers' common questions. That pattern will help you choose one change to tip the scales in your favor, and then you can start testing.

The trouble usually arrives about four weeks later, when the numbers have moved and you don't know why. Was it your change? Did a rival publish something new? Or did the engine simply answer differently that day?

When you ask an AI engine the same question twice, the list of brands rarely matches, so a single before-and-after check can't tell these causes apart. Your own counts can, because they show where to start and whether the change held.

What is AI visibility optimization?

AI visibility optimization is changing your pages, your presence on other sites or your crawler access so AI answers name and link your brand more often.

The result is a count, not a task list: the share of collected answers that name you, and the share that link your site. Firemark's guide to named, linked and recommended mentions keeps those two apart, and so should your baseline.

The craft isn't separate from SEO, at least for Google. Google's guide says SEO best practices continue to be relevant, because its generative AI features are rooted in core Search ranking and quality systems. That statement covers AI Overviews and AI Mode on Google Search. It says nothing about other engines, which is one more reason to measure what happens instead of assuming.

Why measure before you change anything?

A baseline is your first set of counts, taken before you change anything. It's how you tell a real change from the ordinary swing between runs, and without one an edit and a lucky afternoon look the same.

Three cards: a share is shown from 20 answers, a movement is called after 3 comparable runs, and a change of 5 points or more is required.
Firemark method page, read 12 November 2026. The 3 runs and 5 points rule is Firemark's own convention

Ask an AI engine for brand recommendations twice and you'll rarely get the same list. In SparkToro's study, 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI, 2,961 runs in all, in late 2025.

For Google, that meant AI Mode only where no AI Overview showed. For ChatGPT and Google's AI, the chance of the same list of brands in any two responses was less than 1 in 100. Perplexity and Gemini weren't in the study, so read it as a warning about single answers, not as a figure for every engine.

SparkToro's Rand Fishkin concluded that visibility percentage across dozens to hundreds of prompts, run multiple times, is a reasonable metric, while the study leaves open how many runs are enough. That's the baseline: many answers, same questions, several runs.

Change a page first and count afterward, and you have an "after" with no "before", because the answers have already moved on.

Firemark's method page sets its own rules for reading counts. It shows a share only from 20 answers, then draws a 95% range on each engine's own base: the band the share could plausibly sit in, given how many answers sit behind it. NIST's handbook describes the method.

Firemark calls a movement after three comparable runs and a change of 5 points or more. That rule is Firemark's own convention, not a study finding, so use it as a working floor.

How do you read your baseline?

Four patterns in a baseline each point at a different first check: not named, named but not linked, named less than rivals, and named on one engine only.

The table pairs each pattern with the check that costs least and the question that check can't answer.

How do you read your baseline?
Pattern in your countsWhat the counts showFirst thing to checkWhat it can't tell you
Not named in most answersA low named count over the baseWhich sites the answers cite, and whether you appear on themWhether adding a page would change it
Named, rarely linkedNamed count high, linked count lowWhether one page answers the question and can be crawled and indexedWhether the engine links at all for that kind of question
Rivals named more oftenRival counts above yours on the same promptsWhat the rivals' cited pages say that yours don'tWhy the engine prefers them
Named on one engine, not anotherA gap between per-engine countsThat engine's own documentation, then a second runWhether the gap holds across runs

Every count in the table needs its base beside it, such as named in 12 of 40 answers. Firemark's guide to AI share of voice shows how the denominator changes the number.

Pick the row that matches your counts. If two rows match, start with the one that's cheapest to check. A check that costs an afternoon beats a rebuild that costs a quarter.

What should you leave alone?

Google's own guide lists llms.txt files, chunking content, writing just for AI systems and seeking inauthentic mentions as things you can ignore for Google Search. On llms.txt it's direct: Google Search doesn't use the file.

The same guide also says structured data isn't required for generative AI search and that no special schema.org markup is needed, though it remains a good idea for rich results. Those lines say nothing about ChatGPT, Perplexity or Gemini, so a change Google says you don't need may still matter on another engine, and no source cited here says either way.

That gives a plain rule: don't spend a test slot on a change the engine's maker says you don't need. Put the slot on something that engine does document, such as being indexed and eligible to show with a snippet. For the ChatGPT side, Firemark's guide to getting cited by ChatGPT sorts each lever by the evidence behind it.

How do you test one change?

Test one change at a time against your baseline, keep the prompts and engines fixed, and write the change down before the next run.

Four steps: change one thing, keep prompts and engines fixed, log the change before the next run, and give the verdict after repeat runs.
Steps from this article. The three-run rule is Firemark's own convention

Counting a run is covered in Firemark's guide on how to measure AI visibility. What matters here is the log, because a count with no record of what you changed can't be read later. Each row needs the date, the page changed, the lever, the prompts it should affect, the baseline count and base, the count after and its base, the range and your verdict.

Note anything else that moved in the same window: another edit to the same page, a rival publishing a new page, an engine update. Engines update on their own schedule and rivals publish on theirs, and neither sends you a calendar invite. Any of them can explain a shift you were about to credit to your change.

What does a worked example show?

A worked example shows that a 15-point rise can sit inside sampling noise on 40 answers and outside it on 200.

Here's one log row as an illustration, arithmetic only and not a real brand. You rewrite one page and ask the same basket, a fixed set of 40 prompts on one engine, before and after.

Before the change, the brand is named in 8 of 40 answers: 20%, with a 95% range of 11% to 35%. After it, the brand is named in 14 of 40: 35%, with a range of 22% to 50%.

That's a 15-point rise, and you might be thinking the job is done. On 40 answers it isn't, because the two ranges overlap. Firemark's rule would call it a movement once three runs agree, but overlapping ranges only make that call provisional.

So collect more answers. With a basket of 200 prompts, named in 40 of 200 is 20%, range 15% to 26%, and named in 70 of 200 is 35%, range 29% to 42%. Both pairs are arithmetic, not data from a real brand.

Ranges for a brand named in 8 of 40 then 14 of 40 answers overlap; at 40 of 200 then 70 of 200 answers they are 15% to 26% and 29% to 42%, which do not.
95% ranges. Arithmetic only, not data from a real brand

Ranges still overlap? Keep collecting. Ranges apart? Then sampling noise alone is a poor explanation for the rise, and the 200-prompt basket lands there.

The share didn't change between the two examples. The number of prompts did.

Can Search Console check a change?

Search Console can give a direction check on Google's side, though not a match for your counts. Its Generative AI performance report lists impressions for your URLs in AI Overviews and AI Mode, grouped by page, country, device and date. Google rolled it out to all websites worldwide as of August 31, 2026.

It reports real-user impressions for your URLs, not your fixed prompt set, so it can't show whether an answer named your brand without linking to you.

What can one test not show?

One site's test can show whether a count moved after a change, and it can't show why.

A rise after an edit is consistent with the edit working. It's also consistent with a rival losing ground or an engine changing its behavior. A result on one engine isn't a result on another: engines answer differently, and each count has its own base.

Quote: one site's test can show whether a count moved after a change, and it can't show why.

And a count says what answers contain, not what readers did next.

Firemark's method page lists what its own tracker doesn't cover: personalized or logged-in answers, clicks, visits or sales that follow an AI answer, and countries and languages outside US English and the markets added on request. Any tool or sheet you use has limits of its own, and you should write them down next to the numbers.

An unchanged count is a finding too. It says that lever, on that page, didn't move the answers over those runs, which saves the next test from repeating it. A flat result still goes in the log, so the same idea isn't rerun by mistake.

What to tell your boss or client is short: the count, the base, the range, and one sentence on the single change made. If a page was edited and nothing moved, say that plainly.

Keep the log where the whole team can read it, with one row per change. A log that lives in one person's head is a rumor. Fill in the verdict only after the repeat runs, not after the first good one.

Frequently asked questions

Is AI visibility optimization the same as GEO?

The terms overlap in common use. Here, AI visibility optimization means any change you make so AI answers name and link your brand more often, judged by counts over a base.

How long until a change registers?

No source cited here gives a reliable time, so none is offered. Measure on a fixed schedule and judge a change after three comparable runs, which is Firemark's rule. For Google, a page has to be indexed and eligible to show in Search with a snippet before it can appear in generative AI features, and the site must be included in those features in Search Console.

Does llms.txt help AI visibility?

Google says Google Search doesn't use llms.txt, including for its generative AI capabilities. That statement doesn't establish what other engines do, so if you want to know, test it as your one change and log the result.

Do you need a tool to optimize AI visibility?

A small set doesn't need a tool. A sheet, 20 or more collected answers and a fixed schedule are enough to start a baseline by hand, though a base that small will only show large moves. A tool saves the time of asking and recording. Firemark sells one, so weigh this answer with that in mind.

Should you optimize for every AI engine at once?

Counts are drawn per engine, so test one engine's pattern at a time. A change aimed at Google's AI features can be sanity-checked in Search Console. A change aimed at another engine can't, so it needs its own repeat runs.

Measure first, then change one thing

Choosing your first AI visibility optimization comes down to one count over a base, taken before the change and after it. Firemark counts named and linked answers over a base and draws the range, so you can prove it worked. The demo report shows a full report on demo data, and the method page lists what the counts leave out.

Angelique Swain AI-first SEO Strategist

Written by

Angelique Swain · AI-First SEO Strategist

Angelique Swain is an AI-first SEO strategist with over a decade in organic search across retail, medical and B2B software. Most of her work now is on how AI engines choose which brands to name. She builds fixed prompt sets, re-runs them to see how much the answers vary, and traces what the cited pages have in common. At Firemark she writes about measuring AI visibility and what to do once you can see it.

  • Cover reading How to Track AI Traffic in GA4 and What It Misses, in cream type on a dark Firemark ground

    Measure AI visibility

    How to Track AI Traffic in GA4 and What It Leaves Out

    How to track AI referral traffic in GA4 with one custom AI channel that combines Google's AI Assistant channel and a tested regex, plus what GA4's count leaves out and why.

    · 7 min read

  • Cover reading Unlinked Brand Mentions in AI Answers: What They Mean, in cream type on a dark Firemark ground

    Measure AI visibility

    What Do Unlinked Brand Mentions Mean in AI Answers?

    An unlinked brand mention in an AI answer is an answer that names your brand and cites none of your pages. It happens when the answer cites nothing or cites other sites, and only the second points at pages you can work on.

    · 6 min read

Find out if AI recommends you

Start tracking your brand. New accounts get one free run.