All articles
10 min read

AI Visibility Optimization: Prove Each Change With Enough Answers

AI visibility optimization only counts when you can prove it. Change one thing, then compare your brand’s mention rate before and after across roughly 140 to 360 AI answers on each side, because a single answer varies too much to trust.

By Harjot Chopra, Co-Founder, PageLens.ai · Updated 19 Nov 2026

A person at a desk compares two matching sheets of dotted marks side by side to see whether anything changed.

Key takeaways

  • Prove it with a before and after comparison of mention rate. From a 20% starting rate, a 15-point lift (to 35%) needs about 138 answers on each side, and a 10-point lift (to 30%) needs about 294.

  • One answer proves nothing. Repeated runs of the same prompt return different brands, different URLs and a different order, so judge the rate across many answers, not a single check.

  • Judge each engine separately. Engines draw on different sources and favor different brands, so a win on one does not carry over to another.

  • Test one proven-sounding change, not a bundle. Google’s guide says no special files or AI-only rewrites are needed, so pick one change, such as crawlable, original pages or authentic third-party mentions, and run it as a clean test.

In this article
  1. 1What is AI visibility optimization, and how is it different from AI visibility?
  2. 2What is the test that proves AI visibility optimization worked?
  3. 3Why can't you judge AI visibility optimization from one check?
  4. 4How many answers do you need before a change counts as real?
  5. 5How should you read the result of an AI visibility test?
  6. 6Is AI visibility optimization just SEO, and what should you change first?
  7. 7How do we run AI visibility optimization tests at PageLens.ai?
  8. 8What should you do before your next test?
  9. 9Frequently asked questions

AI visibility optimization only counts if you can prove it, and a single check cannot. Ask an AI engine the same buying question 60 to 100 times and nearly every answer differs.

In SparkToro and Gumshoe.ai research that gathered 2,961 responses to 12 prompts across ChatGPT, Claude and Google’s AI search, the same list of brands came back less than 1% of the time, as Search Engine Journal reported on January 30, 2026.

That is how a win turns out to be noise. A team rewrites a page, tests one prompt, sees the brand named and calls it a win. Run the same prompt tomorrow and the answer may differ.

This page gives you the test, the answer counts to aim for, a shortlist of what to change first, a decision table and a log template to record each result. The effort is modest: twenty prompts run seven times each gives 140 answers on each side, and the rest is recording the numbers. For the full list of tactics, see our guide to answer engine optimization.

What is AI visibility optimization, and how is it different from AI visibility?

AI visibility is the outcome: how often AI engines name, cite or recommend your brand when buyers ask relevant questions. AI visibility optimization is the work that tries to raise it.

Answer engine optimization (AEO) and generative engine optimization (GEO) are other names for that work, and Google's guide treats the Google part of it as SEO. Because the outcome varies from run to run, this page treats every change as a hypothesis to test, not a result.

What is the test that proves AI visibility optimization worked?

The test compares your mention rate before and after one change, using the same prompts, the same engine and enough answers. AI visibility optimization means changing your site, or the pages around it, so AI engines name or cite your brand more often. Mention rate is the share of test answers that name your brand.

Run these seven steps. Each has a pass condition.

  1. 1
    Freeze the prompts.

    Pass: 20 to 30 buyer questions saved in a file, with wording locked. Any edit restarts the baseline. Prompt research covers how to choose them.

  2. 2
    Pick one change and one engine to judge first.
    Pass: you can state the change in one sentence, and nothing else on those pages changes during the test.
  3. 3
    Run the baseline to the answer count in the table below.
    Pass: you reached the count, spread over at least three different days (our recommendation, to avoid one odd day setting the number).
  4. 4
    Record the baseline mention rate and its margin of error.
    The margin of error is the range either side of a measured rate within which the true rate probably sits. Pass: both numbers are in the log.
  5. 5
    Ship the change, confirm it was refetched, then wait.
    Our recommendation: wait at least four weeks after the change is live and recrawled, and rerun on the same weekdays as the baseline. Pass: crawler hits for the changed page appear in your server logs (or, for Google, in Search Console crawl stats), and the four weeks have passed.
  6. 6
    Rerun the identical prompts, engine and answer count.
    Pass: same prompt file, same count.
  7. 7
    Apply the decision table.
    Pass: the call is written in the log with its numbers.

Why can't you judge AI visibility optimization from one check?

One check misleads because the same prompt returns different answers almost every time. Three separate measurements point the same way.

  • Brand lists. In the SparkToro and Gumshoe study, the same brand list came back under 1% of the time and the same list in the same order under 0.1%. Four headphone brands (Bose, Sony, Sennheiser, Apple) appeared in 55% to 77% of 994 responses to varied headphone prompts, so share across many runs is steady where list position is not.
  • Cited pages. A test that ran 10,000 keywords through Google AI Mode three times on the same day found that only 9.2% of URLs and 14.7% of domains appeared in all three runs, according to Search Engine Journal's July 1, 2025 report.
  • Named providers. In our Citation Research Study 001 (800 answers, each question asked twice, October 5, 2026, Indian exam questions in English), only 80 of 400 repeat pairs named exactly the same providers, 20%. The first-named provider held in 64% of 374 comparable pairs.

Two conclusions follow. Track the share of answers that name you, not your position in a list. And treat any single before and after pair as noise until the answer count is high enough.

Five identical mugs in a row hold noticeably different pours of coffee from the same pot.

How many answers do you need before a change counts as real?

For a lift from 20% to 35%, you need about 138 answers on each side. For 20% to 30%, about 294. The smaller the lift you want to detect, the more answers you need.

This is our model, not an official figure. It is a two-sided comparison of two proportions at 95% confidence and 80% power. The answer count per side is (1.96 × √(2 × p̄ × (1 − p̄)) + 0.8416 × √(p1 × (1 − p1) + p2 × (1 − p2)))² ÷ (p2 − p1)². Here p1 is the before rate, p2 the after rate and p̄ their average, and we rounded each result up. 95% confidence means the margin is set so that, over many repeated tests, about 95 in 100 would contain the true value. 80% power means a real lift of the size you chose would be detected in about 80 of 100 tests.

Mention rate before → afterAnswers needed on each side
10% → 20%199
20% → 30%294
20% → 35%138
30% → 40%356
30% → 45%163
50% → 65%170

The model treats every answer as independent. Repeated runs of one prompt are not fully independent, so treat these counts as floors. Twenty prompts run seven times each gives 140 answers per side.

To size the margin on a single mention rate, use 1.96 × √(p × (1 − p) ÷ n), where n is the number of answers:

AnswersMargin at a 50% rateMargin at a 20% rate
30±17.9 points±14.3 points
60±12.7 points±10.1 points
100±9.8 points±7.8 points
200±6.9 points±5.5 points
400±4.9 points±3.9 points

A before and after comparison needs the wider margin of the difference (our model): 1.96 × √(p1 × (1 − p1) ÷ n1 + p2 × (1 − p2) ÷ n2), where n1 and n2 are the answer counts before and after. At 140 answers per side, a 20% rate has a single-rate margin of about ±6.6 points, but the margin of the difference between 20% and 35% is about ±10.3 points. At 30 answers on each side, a rise from 40% to 55% is a 15-point move, but the margin of the difference is about ±25 points, so it is not yet a result. Our guide to AI citation tracking goes further on how many answers make a number real.

How should you read the result of an AI visibility test?

Read it against the margin of the difference, per engine, and decide before you look. Use this table.

Result after the second runCallAction
Rate rose by more than the margin of the difference, answer count reachedKeepKeep the change and test it on the next page group
Rate moved by less than the margin of the differenceInconclusiveExtend the run, or stop; do not report a win
Rate fell by more than the margin of the differenceRevertRoll back, then look for what else changed
Rose on one engine, flat on othersEngine-specificReport that engine only

Per-engine reading matters because engines draw on different sources. In Study 001, the four engines agreed on the most-named provider for only 5 of 20 exams.

Study 001 covers Indian exam questions, so treat it as an illustration, not a benchmark. Our inference: a change that suits one engine's sources can do nothing for another's.

Log every test in one row:

DateChangeEnginePromptsAnswers beforeRate beforeAnswers afterRate afterCall

Is AI visibility optimization just SEO, and what should you change first?

Mostly yes, and the part that is new is the measurement. Google states that its generative AI features are rooted in its core ranking and quality systems, so SEO best practice stays relevant. Its optimization guide treats AEO and GEO as SEO and adds that you don't need new machine-readable files (llms.txt is named), special rewrites for AI, or structured data for generative AI search.

Our recommendation: pick your first change from this table, apply it to one page group, then test it with the steps above.

Change to test firstWhat the evidence saysHow to test it
Make key pages crawlable and indexedGoogle says a page must be indexed and eligible to show with a snippet to appear in its AI features

Confirm indexing in Search Console, then run the access test in our guide to answer engine optimization

Publish first-hand, non-commodity content: original data, tests, expert viewsGoogle says unique, useful content will likely influence AI search visibility more over time than any other suggestion in its guideShip one new or rewritten page group and change nothing else
Earn authentic mentions on other sitesGoogle says seeking inauthentic mentions is not as helpful as it seems, so the mentions worth testing are real ones

Run the 20-prompt audit in our guide to third-party mentions first

Add llms.txt, chunk pages or rewrite "for AI"Google says Google Search ignores llms.txt and needs no chunking or AI-only rewritingSkip as a Google tactic. If you add a file for another reason, test it like any other change

Our GEO vs SEO guide covers the three differences between the two and gives a one-hour test.

When you evaluate any dashboard, ask how many answers sit behind each number. A score from a handful of runs is inside its own margin.

How do we run AI visibility optimization tests at PageLens.ai?

We run a baseline audit, then track a fixed prompt set daily on a paid plan, so each before and after comparison uses the same prompts. Our Free AI Visibility Audit takes a domain with no signup or credit card. In about 30 seconds it runs real buyer prompts across several AI engines and returns a visibility score from 0 to 100, what the engines say about you and the competitors they name instead. A single audit run is a starting picture, not a before and after test.

For the test, our plans track prompts daily. As of November 2026 (see pricing):

  • Launch, $299 a month: ChatGPT, Google AI Mode and Perplexity.
  • Growth, $699 a month: adds Gemini and Grok.
  • Enterprise, from $1,499 a month: adds Claude and Copilot, seven engines in all.

Launch and Growth both include 100 visibility prompts a day, across 10 categories and up to 10 competitors each. Even with 100 prompts a day, spread a 138-answer baseline over at least three days to meet step 3, and a 294-answer baseline over about four. Check how your own prompt set maps to the allowance before you start.

Our prompt research feature surfaces the questions your buyers ask AI, which is where the prompt file in step 1 can start. Our analytics feature combines organic traffic, search rankings and the AI citations each published page earns, through a Google Search Console connection.

A marketing professional reviews a steady tracking dashboard at a desk next to a printed checklist.

What should you do before your next test?

Before you change another page, make three decisions: the one change you will test, the engine you will judge first, and the answer count your expected lift needs. From a 20% starting rate, a 15-point lift needs about 138 answers on each side and a 10-point lift needs about 294, spread over several days.

If the second run lands inside the margin of the difference, call it inconclusive and extend the run rather than report a win. Log the result either way, because a flat result tells you which tactic to drop.

In the next 48 hours, freeze your prompt file and start the baseline. If you would rather not collect hundreds of answers by hand, Book a demo and we will help you plan a test on your own prompts.

Run your baseline with a free AI visibility audit

See where ChatGPT and other AI engines name your brand today, with no signup or credit card. Then book a demo to plan a before and after test on your own prompts.
See PageLens pricing

Frequently asked questions

Treat it as a series of single-change experiments: lock a prompt file, take a baseline, ship one change, rerun, then keep or revert it. Copy a change to more pages only after it passes the decision table.
Track 20 to 30 buyer prompts, and run each repeatedly until you reach the answer count your expected lift needs.
Wait at least four weeks after the changed page is live and recrawled, then rerun on the same weekdays as the baseline. That is our recommendation, not a measured figure. Confirm the refetch in server logs or Search Console crawl stats.
Trust it when you know how many answers produced it and the change you are reading is larger than the margin of the difference. A score from a single run of a few prompts is a snapshot, not a trend.
Google says no new machine-readable files are needed for its generative AI search, and it names llms.txt, because Google Search itself doesn't use them. Test any file you add the same way as any other change.

Sources

  1. 1.Search Engine Journal, "AI Recommendations Change With Nearly Every Query: SparkToro" (January 30, 2026)
  2. 2.Search Engine Journal, report on a Google AI Mode consistency study (July 1, 2025)
  3. 3.Google Search Central, "Optimizing your website for generative AI features on Google Search"
  4. 4.Citation Research Study 001: AI coaching recommendations (fieldwork October 2026)
  5. 5.PageLens.ai pricing (as of November 2026)
  6. 6.PageLens.ai Free AI Visibility Audit
  7. 7.PageLens.ai prompt research feature
  8. 8.PageLens.ai analytics feature

Keep reading

PageLens.ai.

Measure how AI engines see your brand, then turn the gaps into growth.

© 2026 PageLens.ai

Powered by PageLens.ai

Discover how often AI recommends your brand.