AI visibility: how to measure and improve it

Why is AI visibility a different metric from search rankings?

Because ranking well no longer guarantees being cited. Ahrefs analysed 863,000 SERPs and 4 million AI Overview URLs in a March 2026 study and found that only 37.9% of URLs cited in AI Overviews also ranked in the organic top 10 for the same query. Around six in ten cited sources came from somewhere other than the visible top ten.

The mechanism behind that gap is query fan-out. Google states that AI Mode "uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf" (Google, May 2025). You are no longer competing for one keyword. You are competing across a spread of sub-questions the user never typed and never sees.

The second difference is what a win looks like. A ranking win is a click. An AI win is often a mention with no click behind it. Pew Research Center tracked 68,879 Google searches from 900 US adults and found that when an AI summary appeared, people clicked a traditional search result in 8% of visits, against 15% of visits without a summary. Clicks on a link inside the summary happened in 1% of visits.

Traffic is not the only payoff either. Adobe data reported in June 2026 found AI referrals to US travel sites grew 194% year over year in May 2026, and those visitors spent 70% longer per visit with bounce rates 41% lower than non-AI traffic, although they still converted 28% less. That figure is travel-specific, so treat it as a signal about visitor behaviour rather than a number for your own vertical.

For the underlying mechanics, see how AI visibility differs from classic SEO.

Why can’t Search Console or analytics tell you this?

Google added a generative AI performance report to Search Console in 2026. It covers impressions from AI Overviews and AI Mode, grouped by page, country, date and device. It is genuinely useful, and it is also bounded: impressions only, and only Google’s own surfaces.

Nothing in that report sees ChatGPT, Perplexity or Claude. Analytics sees only the people who arrive, and most AI answers end without a visit. If you want to know whether an engine recommends you, you have to ask the engine.

How is AI visibility actually measured?

By sampling, not by counting. A tracker defines a set of prompts a real buyer might type, sends them to each engine on a schedule, then parses every answer for brand mentions, cited URLs and tone. That is the entire method, and its credibility depends on being upfront about what it is.

What is a good AI visibility score?

There is no published industry benchmark, and any vendor quoting a universal "good" number is guessing. Scores depend on how broad your prompt set is, how crowded your category is and how old your brand is, so two companies with identical scores can be in completely different positions.

Use three internal reference points instead. First, your own trend: is the score moving up across a stable prompt set. Second, relative position: your Share of Voice against the set of names that actually appear alongside you. Third, coverage: the count of tracked prompts where you never appear at all, which is usually the most actionable number on the page.

For a live reference point that needs no account, the public visibility leaderboard shows how brands currently rank in AI answers.

Why is your brand invisible in AI answers?

Usually one of five reasons, and they are diagnosable.

• Your pages answer a keyword rather than a question, so no self-contained passage matches the sub-queries an engine actually issues.
• The answer exists but is buried below 800 words of preamble, past the point where a retrieval step would extract it.
• The content is unciteable rather than invisible: no dates, no numbers, no named sources, nothing a model can safely repeat.
• The problem is off your site entirely, because the engine leans on third-party pages that discuss your category without you.
• It is technical, and the substance lives in JavaScript, PDFs or images that never resolve into crawlable text.

What actually moves AI visibility?

The strongest evidence is a peer-reviewed paper. Aggarwal and colleagues, in "GEO: Generative Engine Optimization" (KDD 2024), tested content edits across a 10,000-query benchmark and found that adding direct quotations from credible sources raised a source’s visibility by roughly 40% on their headline metric, with adding statistics and explicit citations both worth around 30%. Keyword stuffing performed worse than the untouched baseline.

Off-site signals appear to matter too, though the evidence is weaker. Ahrefs studied 75,000 brands in December 2025 and found brand mentions on YouTube correlated with AI citation at about 0.74 and branded web mentions at roughly 0.66 to 0.71, while Domain Rating sat between 0.27 and 0.33. The authors are explicit that correlation is not causation, and big brands accumulate both mentions and citations for the same underlying reason. Read it as a lean toward earning real brand presence, not as a lever with a guaranteed return.

The practical priority order that follows: put a liftable answer near the top of every page, attribute your facts to named sources with dates, cover the sub-questions rather than the head keyword, and build genuine mentions on the platforms your buyers already read. That is the core of generative engine optimization.

How does AI visibility differ from engine to engine?

Substantially, which is why a single blended score hides more than it shows. An analysis of more than 680 million citations collected between August 2024 and June 2025 found each engine leans on a different source mix.

Two practical consequences. A brand can be well established in one engine and absent from another, so per-engine breakdowns matter more than the average. And these mixes shift with platform deals, licensing agreements and model updates, so any engine-specific tactic needs re-checking rather than being set once.

How often should you measure AI visibility?

Weekly for the tracked prompt set, monthly for the analysis. Weekly runs give you enough samples to average out the non-determinism described above. Monthly is the honest cadence for drawing conclusions, because content changes need time to be crawled, indexed and picked up by retrieval.

Add an off-cycle run after two events: a significant content push on your side, and a visible model or engine update. Comparing the same frozen prompt set before and after is the closest thing to a controlled test available in this channel.

How do you turn AI visibility data into action?

A score on its own changes nothing. The loop that works is measure, find the gap, publish against it, then measure the same prompts again.

Optimization for AI runs that loop end to end. It tracks your domain across ChatGPT, Perplexity, Google AI Overviews, Gemini and Claude, alongside your classic Google organic position, and computes Visibility Score and Share of Voice from the prompt set you define. It then surfaces the prompts where you are missing, generates and publishes articles built to answer those specific questions, and re-measures the same set. You can see how the measurement works in detail.

The point of the closed loop is falsifiability. If a published page does not move the prompts it was written for, you learn that within a measurement cycle instead of assuming it worked.