How to Measure AI Search Visibility

OpenAIAnthropic ClaudeGoogle AI SearchPerplexity
Ask AI →

Your AI visibility score went up four points last month. Nobody on the team can say why. Next month it will drop, and the vendor will ship a new dashboard that explains the drop after the fact.

I have watched this play out at four companies this year. Same pattern, different logos.

TL;DR

  • The number: A vendor score reflects one prompt set on one day, so treat it as a smoke test, not a measurement of your market.
  • The system: Three layers do the actual work, share of answers, citation quality, and branded demand.
  • The scale: AI Overviews reached 82.91% of US desktop branded searches by late September 2026, per Ahrefs.
  • The gap: 45% of marketing leaders cannot measure their brand inside AI answers at all, per Semrush.
82.91%
of US desktop branded searches showed an AI Overview on September 29, 2026 (Ahrefs)
45%
of marketing leaders cannot accurately measure brand visibility in AI answers (Semrush)
15 vs 3
sources cited per answer by ChatGPT versus Gemini, a 5x difference in citation surface (Semrush)

The Score Is Not the Measurement

Every AI visibility tool sells a single number. It arrives with a trend line and a competitive rank. It feels like a metric you can manage, and that is exactly the problem.

I have bought two of them. Both were wrong, in opposite directions.

Look at what the underlying data actually did this year. Ahrefs tracked 232.8 million US desktop searches from July through September 2026. AI Overviews appeared on 61.3% of branded keywords in July, 65.1% in August, and 73.3% in September. By September 29 the daily figure hit 82.91%. The same study found that 76.2% of branded AI Overviews sat in position one, above the brand’s own homepage.

Now read the click data from Seer Interactive. Across 53 brands and 5.47 million queries, organic click-through rate on AI Overview queries fell to a floor of 1.3% in December 2025. Then it rebounded to 2.4% by February 2026. The agency’s own projection of continued decline did not hold.

One dataset says coverage is expanding fast. Another says the click damage paused. A third, SparkToro’s panel work, puts zero-click Google searches at 68.01% in early 2026. Three credible sources, three different directions, one conclusion for an operator. A single composite score cannot hold all of that, so stop asking it to.

Vendors sample different prompts, different platforms, and different days. Two tools can both be right and still disagree by thirty points. That is not a contradiction. It is a reminder that the score is a sample, and you do not control the sample.

The operating rule

Measure the trend you can explain, on a sample you control. A number you cannot decompose is a mood, not a metric.

LinkedIn Algorithm Report 2026

Layer 1: Share of Answers on a Fixed Panel

Start with a prompt panel you own. It is forty to sixty questions your buyers actually type, lifted from call notes, chat logs, and sales objections. Write them down, freeze the wording, and never let the vendor swap them for you.

I keep the list under sixty prompts. Forty is enough.

1
Harvest the prompts from real conversations

Pull the first question from your last twenty discovery calls. Add the comparison questions buyers ask when they shortlist. Those are your prompts, not the ones a tool invented.

2
Lock the wording for a quarter

Changing a prompt mid-quarter resets your baseline. Version the list and date every change so a shift in the number has a known cause.

3
Run the same list, same day, every week

ChatGPT, Gemini, Google AI Mode, and AI Overviews on the same afternoon. Same order, same phrasing, no personalization where the platform allows a clean session.

4
Record three states, not a score

Mentioned, cited, or absent. Mentioned means your name appeared. Cited means your domain was linked. Absent means neither. A percentage of forty prompts is honest. A weighted index is not.

Two columns matter more than the total. Mentioned-but-not-cited tells you the model knows you and will not vouch for you. Cited-without-mention tells you an owned page earned the link but the brand story did not land. Both are fixable, and neither shows up in an aggregate score.

Expect the first month to look bad. Models change quietly, and a tight panel catches shifts that a monthly report smooths over. Log every platform update next to that week’s result.

Layer 2: Citation Quality by Platform

Platforms do not cite the same way, so one blended number hides the whole picture. Semrush analyzed 126 million US AI search prompts for its 2026 index. ChatGPT cites an average of fifteen sources per response. Gemini cites an average of three.

PlatformSources per answerWhat it leans onWhat to track
ChatGPT~15Reddit, Wikipedia, community threadsWhich URL gets cited, and whether it is yours
Gemini~3Wikipedia, Reddit, YouTubeMention-to-citation gap
AI OverviewsVaries by queryWikipedia at 71.9% of external citationsPosition of the overview versus your page

Semrush also found that on Gemini, the overlap between brands mentioned in an answer and domains cited as sources can be as low as 30%. That gap is the whole game. A mention without a citation is a rumor the model repeats. A citation is a link a buyer can click.

I still check this by hand. Two prompts a week is enough to catch a change.

The Ahrefs brand panel makes the same point from the other side. Among external citations in branded AI Overviews, Wikipedia accounted for 71.9%, YouTube 38.6%, and LinkedIn 22.8%. Third parties are doing the vouching, and most brands have no plan for the pages doing it.

So track citation sources per platform, not a global visibility rank. Ask a simpler question each week. Which domains are being used to describe us, and do we influence any of them? The answer usually names a Wikipedia entry, two review sites, a podcast transcript, and a competitor’s comparison page. If you want the content side of that fix, I wrote about what actually gets cited earlier this month.

Then fix the cheapest ones. Correct the profile fields. Publish the comparison page you have been avoiding. Give the analyst a briefing document with the numbers you want quoted.

Layer 3: Branded Demand Is the Outcome

Visibility is an input. Demand is the result. If AI answers are shaping the shortlist, that shows up in the searches people run after they read the answer. That is why the budget conversation has to reach this layer. Which line item you fund matters more than which tool you renew. Where the AI SEO money goes is the part most teams skip.

I trust this layer more than the other two. It is harder to fake.

Pull branded impressions and clicks from Search Console. Pull direct sessions from analytics. Pull branded search volume from your keyword tool and inbound demo requests that name a brand term in the form field. Four sources, one weekly line, and a direction of travel that survives a platform update.

Your CFO does not need the prompt detail. Give them one chart with three lines and one sentence of interpretation. Share of answers, citation quality, branded demand. If a line moves and the sentence cannot explain it, the chart is not ready. This is the same discipline behind the marketing scorecard that lies to most teams, applied to a newer channel.

This layer is slower than the others, and that is the point. A rise in share of answers with no move in branded demand is a content problem, not a visibility win. A rise in branded demand with flat citations means the model is not doing the work, and you are paying for it elsewhere.

Read the three layers together and the diagnosis writes itself.

What I Actually Think About AI Visibility Scores

Every time I audit a marketing operation, the first thing I ask for is the prompt list behind the AI visibility score. Most teams cannot produce it. They have a number, a trend, and no idea what was asked.

Koka Sexton
Koka Sexton
B2B Marketing · Revenue Architecture
1h ago

A visibility score is a poll of a machine. It tells you what one model said, on one day, with one prompt set. Treat it like a smoke test. Never treat it like a P&L line.

189 Likes · 43 Comments

The pattern I see is ownership. When one named person owns the prompt panel, the numbers stay consistent and the weekly review takes twenty minutes. When the dashboard belongs to everyone, nobody reads it, and the only time it gets opened is the month a score drops.

My second take is less popular. Most teams should not buy another visibility tool this year. They should buy nothing and run the manual sample for a quarter. Forty prompts, four platforms, one spreadsheet, one owner. You will learn more in eight weeks than a year of vendor reports will teach you. The data will be yours, and you will know what every row means.

That advice costs a tool vendor a renewal. It also keeps a marketing team from confusing a scoreboard with a strategy.

I know how that sounds from someone who sells advisory work. Run the manual sample first, then decide what to buy.

The 30-Minute Weekly Review

The measurement only works if it fits inside a normal week.

Monday, run the panel and log the three states. Wednesday, check citations for the two prompts that moved most, and log which domain got the link. Friday, write one line for leadership: share of answers, citation source that changed, branded demand direction. No deck, no dashboard tour, no forty-slide QBR.

Once a quarter, re-cut the prompt list against fresh call notes and version it.

If the numbers move and you cannot explain the move in one sentence, the sample is wrong. Fix the sample before you fix the content. That rule alone will save a quarter of wasted work.

Key takeaway

AI search visibility is an operations problem dressed as an SEO problem. Own the prompt panel, score three layers, and give the number to one person. That person explains it every Friday, in one sentence.

“Life favors the specific ask and punishes the vague wish” is the line from Tools of Titans. It applies to the questions you put to a model as much as the ones you put to your board.

If you want a second set of eyes on your prompt panel and the three-layer readout, book a strategy call and bring the list. I will tell you which two prompts to fix first.

ways to help 3

About Koka Sexton

Koka Sexton is a marketing leader, strategist, and creator known for pioneering social selling and modern demand generation. With a background spanning startups and global brands like LinkedIn and Slack, he specializes in turning marketing programs into measurable growth engines. A U.S. Army veteran and lifelong builder, Koka combines structure, creativity, and AI innovation to help companies drive scalable revenue impact.

Ways I Can Help

I work with founders, marketing leaders, and growth teams to build smarter, faster go-to-market systems that drive measurable results.

Core Services

  • Go-to-Market & Demand Generation: Develop data-driven strategies that expand pipeline and accelerate revenue.
  • Custom GPTs for marketing: Leverage custom AI agents for marketing tasks to improve campaigns and launch projects faster.
  • Marketing Operations & Automation: Implement AI-enhanced workflows, CRM systems, and marketing tech stacks to optimize performance.
  • Social & Community Strategy: Leverage social selling, influencer engagement, and community platforms to strengthen customer relationships.

More Articles & Posts