How to Measure LLM Visibility: Citation Share, Not One Screenshot

To measure LLM visibility, stop treating an AI answer like a ranking and start treating it like a poll. Fix a written set of 30 to 50 questions your buyers ask, run each one five times per engine in a clean browser, and log which domains get linked. Report citation rate and share per engine, with an error range. Then check that sample against the census Google now publishes: the Search Console Generative AI performance report, live for every site since 31 August 2026. It is for marketing leads and owners who need a defensible number, and it settles one decision: change pages only when a move beats the noise.

Scale is why this matters now. Google said on 19 May 2026 that AI Mode had "surpassed one billion monthly users" (Google I/O 2026 Search announcement). A rank tracker built for ten blue links sees none of those answers.

What changed in Google's AI Overviews and AI Mode reporting?

Google now gives every site owner an official count of AI visibility. The Search Generative AI performance report, announced on 3 June 2026 and rolled out to all websites worldwide on 31 August 2026, shows how often links to your site appeared in AI Overviews and AI Mode. It counts impressions only: no queries, no clicks, no competitors.

Three changes landed close together, and each one matters for measurement.

The report itself. Google's Search Central announcement, 3 June 2026 describes "dedicated views of your impressions within generative AI features on Search, such as AI Overviews and AI Mode," plus a separate report for Discover. The Search Console help page for the report defines an impression as how many times links to your site were shown in a generative AI feature. It lists four ways to slice the data (pages, countries, dates, devices) and no query breakdown. Clicks are not separated either: Google's documentation says sites appearing in AI features "are included in the overall search traffic in Search Console." On 24 September 2026 Google added a multimodal filter, so you can now separate searches started from Google Lens, Circle to Search and image uploads (Search Central, 24 September 2026).

The off switch. The Search generative AI control reached all sites on the same date, 31 August 2026. Inclusion is the default. Exclude your site and "You won't receive any traffic or impressions from these features." Child properties inherit their parent's setting, which is a quiet trap: a subfolder property can show zero AI impressions because someone set the domain to exclude.

Personalisation. People can mark Preferred Sources, publishers they want to see more often in "Top Stories, AI Overviews, and AI Mode." On 20 August 2026 Google released a Preferred Sources button that publishers can embed, so a reader can add a site in one click. By then people had already selected more than 600,000 unique sources (Google, 20 August 2026). Your own marketing team is a likely audience to have marked your site, which flatters any screenshot taken from their accounts.

Here is what each instrument can and cannot tell you.

Neither column answers "are we winning?" on its own. Search Console confirms you appeared, but not for which question or against whom. Your panel shows the competitors and the wording, yet only for the 40 questions you picked. The method below uses each to check the other.

Why is one screenshot not a measurement?

A single AI answer is one draw from a distribution, not a fixed result. Google's own documentation says "AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary," and that AI Overviews "often don't trigger" (Google Search Central's AI features documentation). A screenshot captures one draw.

Look at the mechanics. Google describes "query fan-out" as "a set of concurrent, related queries generated by the model" to fetch more results. Each answer is assembled from several hidden searches, so small shifts in any of them change the links. The model changes too. Google made Gemini 3.5 Flash the default in AI Mode for everyone in May 2026, then on 2 September 2026 made Gemini 3.8 Flash available in AI Mode to Google AI Pro and Ultra subscribers (Google, 2 September 2026). A paying subscriber on your team and a logged-out buyer may not be asking the same model.

The arithmetic is blunt. One cited answer out of one run is consistent with a true citation rate anywhere from about 21% to 100% (95% Wilson interval, a standard way to put an error range on a yes/no rate). Three citations in five runs still spans roughly 23% to 88%. That is why even the researchers who introduced generative engine optimisation (GEO, the practice of shaping content so AI answers cite it) ran every experiment "on five different random seeds" and averaged them (Aggarwal et al., GEO, KDD 2024). If a controlled lab averages five runs, a marketing team should not report one.

How to measure LLM visibility without a vendor score

The method is a fixed prompt panel run repeatedly, scored per engine with error ranges, then read against Google's own census. Forty prompts, five runs, one to three engines, monthly: about 200 observations per engine per month, which is enough to see a real ten-point move and cheap enough to run with a spreadsheet.

Build the prompt set once. Write 40 questions in four groups. Category questions ("how do I measure LLM visibility") and problem questions come first, then comparison and brand questions. Use your buyers' words from sales calls and search terms, not your own taxonomy. Freeze the wording. Changing a prompt mid-period resets its history.

Keep engines apart. Google's AI Overviews and AI Mode already differ from each other, and ChatGPT and Perplexity differ more. A 2026 preprint (a paper not yet peer reviewed) that ran 602 controlled prompts found Google's AI answers cited 12.06 sources per prompt on average, against 6.88 for ChatGPT and 16.35 for Perplexity (arXiv preprint 2604.25707, April 2026). Being one link in twelve is not the same achievement as being one in seven, so never blend engines into one score.

Run clean, on the real surface. Use a fresh, logged-out browser profile, a fixed location and the same week each month. Not the CMO's laptop. And measure the product your buyers use: a model's developer API is not the consumer answer, which layers its own search and personalisation on top.

Log five things per run: prompt ID, engine, linked domains in order, unlinked brand mentions, and any factual error about you. That last column is the one most teams skip and the one compliance cares about.

Budget the runs. Forty prompts times five runs times three engines is 600 answers a month. At roughly a minute per answer to run and log by hand (our planning estimate, not a benchmark), that is about ten hours a month, the point where a small script or a disclosed tool starts to pay for itself. Precision costs more than people expect. The error range shrinks with the square root of the run count, so spotting a five-point move instead of a ten-point one takes about four times as many runs.

These are the metrics worth reporting, and the trap in each.

Illustrative example: a firm runs 40 prompts five times a month in Google AI Mode, so 200 runs. In month one it is cited in 30 runs: a 15% citation rate, with a 95% range of about 11% to 21%. In month two it is cited in 38 runs: 19%, range 14% to 25%. The ranges overlap heavily, so nothing has been proven, and the team does not celebrate or rewrite anything. After rewriting eight pages with sourced figures, month three shows 52 runs: 26%, range 20% to 33%. That rate clears month two's upper bound of 25%, though the two ranges still overlap a little, and the census shows impressions rising on the same eight pages. Now the change is worth reporting as a likely gain, not a proof. (Runs of the same prompt are not fully independent, so treat these ranges as the narrowest honest version.)

What actually moves citation share?

Evidence-dense, original pages move citation share; markup tricks and keyword stuffing do not. In the KDD 2024 GEO benchmark, adding cited sources, quotations and statistics lifted visibility 30% to 40% on its main metric, while keyword stuffing offered "little to no improvement." Google's own guide reaches a similar place by a different route. For the page work itself, our marketing consultancy brand leapbuzz offers AI visibility and GEO services.

The GEO paper is the strongest primary evidence available, and it is worth reading closely rather than quoting as "up to 40%." The lift was measured on GEO-bench, a research benchmark answered by GPT-3.5 Turbo rather than by Google, and the authors found it "varies across domains." On Perplexity.ai, tested with 200 sampled queries and the source text supplied as uploads, quotation addition improved their position-adjusted word count measure by 22%. The finding most summaries miss: gains were largest for pages that ranked low in classic search. When every source in the test was optimised, citing sources lifted visibility 115.1% for the fifth-ranked site, while the top-ranked site lost 30.3% on average.

Citation share is relative. Your competitor's better page can shrink your number while your page stands still, which is another reason to log every domain in the answer, competitors included.

Google's position, in its guide Optimizing your website for generative AI features, updated July 2026, is that "optimizing for generative AI search is optimizing for the search experience, and thus still SEO." It tells site owners to ignore llms.txt files, content "chunking" and rewriting just for AI. It also warns that spinning up pages for every fan-out query "primarily to manipulate rankings or generative AI responses" violates its scaled content abuse spam policy.

So where do they disagree? On intent, not substance. The GEO authors rewrite for the machine; Google says write for the reader and the machine will follow. Both reward evidence a reader can check. Add it because a buyer needs it, then let the panel tell you whether citation rate moved. That is the honest version of GEO, and it is the same line our earlier piece on agent search optimization drew: correctness first, markup second.

Which mistakes make AI visibility numbers worthless?

Most AI visibility numbers fail on method, not on effort. The common faults are measuring from a personalised account, blending engines into one score, counting unlinked mentions as citations, trusting vendor scores that hide their prompts, and reading a zero in Search Console without checking the Search generative AI control.

Trusting a black-box score. Google's guide is unusually direct: "Be wary of third-party tools that promise ranking success or claim to use 'internal' Google metrics. No third-party tool has access to our internal ranking or AI systems." A tool can still save hours of running prompts. Ask it for the prompt list and the run count, then for the error range on each engine. If it cannot show them, the score is a mood.

Confusing a sample with a census. Microsoft's AI Performance report in Bing Webmaster Tools, previewed on 10 February 2026, counts citations across Copilot and Bing's AI summaries and shows grounding queries, but says "the data shown represents a sample of overall citation activity." Read each platform's report for what it claims to be.

Reading position. In the main Search Console performance report, an AI Overview "occupies a single position in search results, and all links in the AI Overview are assigned that same position" (Search Console help). Average position cannot tell you whether you were the first source or the eleventh.

Moving the goalposts. Editing prompts, adding engines or switching location mid-quarter breaks the time series. Re-freeze on a schedule.

The work splits by role. The MD or owner should ask for one number: citation share on buying questions, with its range, quarter on quarter. The CTO or technical lead owns the plumbing, meaning a sheet or small script with fixed fields and stored answers, run from a clean profile, plus a check of the AI control on every Search Console property. Compliance gets the accuracy column. An AI answer that misstates your fees or eligibility is a customer-facing claim about your firm that you did not write and cannot edit, so log it and fix the source page it drew from. Our piece on what changes when the buyer is a machine covers that source-of-truth work.

Where Origin Pi stands

AI answer visibility is a measurement problem before it is a content problem. Teams that skip the measurement end up rewriting pages to chase noise, then crediting the rewrite when the noise swings back.

We design visibility measurement the way we design any agent workflow: a fixed input set, repeated runs, a log of every answer and a human who decides. Is this answer accurate? Is this move real? Those calls stay with a person who can see the audit trail of what each engine said and when. We pair that sample with Google's own census rather than a vendor index, because the census is the only full count you will get, and the panel is the only place competitors and wording show up.

Our view on content is unfashionable. Write for the buyer and cite primary sources. Then keep the page that holds those facts correct, because that page is what an engine quotes, and it is what the GEO evidence rewards. Our marketing AI work builds on that.

Your first step this week: open the Search Console Generative AI performance report and check the Search generative AI control on every property. Then write down the 40 questions you want to be cited for, and run each one five times before you change a single page.