Why does my AI visibility score change even when my citations stay flat?

A visibility score usually depends on mentions and on the competitor set, not only on citations. If a competitor in the locked set gains mentions, your share of voice falls while your own citations hold. A model update can also change how an engine composes answers. On a small prompt set, run-to-run noise alone can move the score. Check the denominator first, then model changes, then noise. Report a change only when it holds across consecutive runs.

Does a CTR collapse mean I lost visibility or just lost clicks?

Not on its own. Read Search Console impressions for the affected queries together with whether an AI answer appears and names your brand. Stable impressions and stable AI presence with fewer clicks mean you lost clicks, not presence: users read the answer and did not visit. Falling impressions, or your brand missing from the answer, mean you lost visibility. Google reports AI Overviews and AI Mode traffic inside the Web search type, with no separate AI filter.

How do I track brand mentions in Perplexity for a client KPI?

Lock a set of prompts your client's buyers actually ask, and a fixed competitor set. Run every prompt several times under recorded conditions: logged out, with a stated country and language. Count the answers that name the brand for mention rate and the answers that link to its domain for citation rate, and keep them as separate lines. Date and archive each run, and report a change only when it holds across consecutive runs.

What signals should I measure for AEO/GEO beyond raw citation counts?

Measure mention rate, the share of answers that name the brand, next to citation rate. Add share of voice against a locked competitor set, answer accuracy, meaning whether the engine states correct facts about the brand, and sentiment. On the outcome side, track AI referral traffic through a custom channel group in Google Analytics, plus a self-reported field asking how people heard about the brand, which catches influence that analytics records as direct traffic.

How often should AI visibility be reported to clients?

Monthly suits most clients. Leadership gets a short summary with trend lines per engine, while marketing leads get citation rate and accuracy issues. Position and self-reported attribution can move to a quarterly view. Behind each report sit several runs of the same locked prompt set, because a single run is a sample, not a trend. Flag accuracy errors as soon as they are found, since each one leads to a direct correction.

AI Search Visibility Reporting: The Metrics Clients Need in 2026

The AI search visibility reporting metrics clients need are mention rate, citation rate, share of voice against a fixed competitor set, answer accuracy and AI referral traffic, each reported per engine. Rankings and clicks understate AI influence, because an answer can name a brand and send no visit. Report a change only once repeated runs confirm it.

What AI search visibility reporting measures that an SEO report does not

A ranking report assumes a stable position on a fixed results page. AI answers have neither: each one is composed per query and changes with the engine, the phrasing and the run.

Two signals carry the visibility part of the report. A mention is an answer that names the brand in its text. A citation is an answer that links to the brand's domain as a source. A brand can have one without the other; see how an AI mention differs from an AI citation.

Traffic understates this influence. Pew Research Center, in a study published on 22 July 2025, found that Google users are less likely to click on links when an AI summary appears: those who encountered one clicked a traditional search result link in 8% of all visits. A brand named in that summary was seen, yet analytics records nothing.

For AI Overviews and AI Mode, Google's Search Central documentation, verified on 3 November 2026, says this traffic sits in Search Console: "reported on in the Performance report, within the 'Web' search type", with no separate AI filter described. ChatGPT, Perplexity, Gemini and Claude offer brands no comparable report, so their data comes from your own prompt sampling. The report keeps the two as separate blocks.

Personalisation, location and logged-in state change answers, so the report also states how prompts were run: logged out or through an API, which country, which language.

The core AI visibility metrics, defined once and compared

Visibility metrics: mention rate, citation rate, share of voice, position

Mention rate is the share of tracked answers that name the brand. Citation rate is the share that link to the brand's domain. Keep them as two lines, because engines show sources differently. Perplexity displays its sources alongside the answer. Anthropic's web search tool documentation, verified on 3 November 2026, states that "the response includes citations for sources drawn from search results", which describes Claude through the API.

Share of voice is the brand's mentions divided by all mentions of the brand and its competitor set. Lock the competitor set at the baseline. Adding or dropping a competitor changes the denominator and moves the share with no change in the brand's own presence.

Position, the brand's order among the brands an answer names, means something only over many mentions. Averaged over a handful, it is noise.

Quality metrics: answer accuracy and sentiment

Answer accuracy asks whether what the engine says about the brand is correct: pricing, features, locations, ownership. Score each answer that mentions the brand as accurate, partly inaccurate or wrong, and log each error with its prompt and engine. Sentiment asks whether the framing is positive, neutral or negative.

Outcome metrics: AI referral traffic and self-reported attribution

AI referral traffic is the sessions that analytics attributes to an AI engine. Self-reported attribution is the share of leads who name an AI assistant when asked how they heard about the brand.

Metric	Formula	Question it answers	Audience	Cadence
Mention rate	Answers naming the brand divided by tracked answers	Do engines name us?	Leadership, marketing	Monthly
Citation rate	Answers linking to the brand's domain divided by tracked answers	Do engines use our pages as sources?	Marketing, operators	Monthly
Share of voice	Brand mentions divided by all mentions in the locked competitor set	Are we named more or less than competitors?	Leadership, marketing	Monthly
Position in answer	Average order among brands named, over many mentions	Are we named first or last?	Operators	Quarterly
Answer accuracy	Accurate answers divided by answers mentioning the brand	Is what engines say about us correct?	Marketing, operators	Monthly
Sentiment	Shares of positive, neutral and negative answers mentioning the brand	How are we framed?	Marketing	Monthly
AI referral traffic	Sessions in the AI channel group	Do AI answers send visits?	Leadership, marketing	Monthly
Self-reported attribution	Leads naming an AI assistant divided by leads answering the field	Do buyers say AI influenced them?	Leadership	Quarterly

Report each metric per engine. ChatGPT, Perplexity, Gemini and Claude retrieve and display sources in different ways, so one averaged score hides which engine moved. If a client wants one figure, weight it and print the weights.

In a small-team report, deprioritise three things: a composite visibility score whose formula nobody can explain, an average position computed over very few mentions, and raw mention counts that ignore how many prompts were run.

How many prompts and runs before a change is real

AI answers are not deterministic. The same prompt can return a different list of brands from one run to the next, on the same engine and the same day. A single run is a sample, never a trend.

Hypothetical example, for illustration only, with no measured data. Suppose a brand's true mention rate on one engine is around half. Each tracked answer then behaves roughly like a coin toss, and binomial reasoning says the typical swing of the observed rate shrinks with the square root of the number of answers. On a couple of dozen answers, chance alone can move the observed rate enough to look like a gain or a loss. Quadruple the number of answers and that band only halves.

Report a change as real only when it holds across consecutive runs and exceeds the noise band for that sample size. Otherwise, report it as stable, with the range observed. Run the baseline several times before the first report, so the band is known.

When a visibility score moves while citations stay flat, check three causes in this order. First, the denominator: a competitor in the locked set gained mentions, so share of voice fell although the brand's own presence held. Second, a model update changed how an engine composes answers or selects sources. Third, sampling noise: the movement sits inside the band.

Where AI traffic shows up in analytics, and where it does not

A mention or a citation fires no event in analytics or in the CRM. Only a click does, and only when the engine passes a referrer or a tagged link. Visits from apps that strip the referrer arrive as direct traffic.

To isolate visits that carry a referrer, create a custom channel group in Google Analytics. The help page, verified on 3 November 2026, explains: "You create a channel group by copying an existing group and making modifications." Add an AI channel matching the referrer domains of the engines you track, placed above the default referral channel. For the full setup, see how to set up an AI referral channel in Google Analytics.

Read a drop in click-through rate with care. Seer Interactive's study, dated 4 November 2025, reports organic CTR for queries showing an AI Overview at 1.76% in June 2024 and 0.61% in September 2025, an overall decline of 61%. A CTR collapse alone does not mean lost visibility: read Search Console impressions for the same queries together with whether an AI answer names the brand. Stable impressions and AI presence with fewer clicks means lost clicks, not lost presence. Falling impressions, or a brand missing from the answer, mean lost visibility.

Add a "How did you hear about us?" field to forms and sales calls, with an AI assistant option or free text. It catches influence that analytics records as direct traffic or branded search.

Which metrics go to which audience, and the minimum viable report

Leadership sees a few trend lines and business outcomes, not raw prompt logs: mention rate and share of voice per engine, AI referral traffic and self-reported attribution. Marketing leads see citation rate and accuracy issues, because those turn into content and PR work. Operators see prompt-level detail: which competitor was named instead.

A report outline that holds up in a budget review:

A one-page summary: what moved, what held stable, and what happens next.
Visibility per engine: mention rate and citation rate.
Share of voice against the locked competitor set.
Accuracy issues found, each with its prompt, engine and the correction requested.
AI referral traffic and self-reported attribution.
Next actions, each with an owner.

The companion article shows how to build a monthly client report for AI visibility, section by section.

For a small team, the minimum viable report is one locked prompt set, two or three engines, a monthly report and three metrics: mention rate, share of voice and AI referral traffic. Add answer accuracy as soon as an engine gets a fact wrong.

Set targets against the client's own baseline and competitor set: as of 3 November 2026, we know of no published industry benchmark for these metrics.

Volatility needs a plain sentence for non-technical stakeholders. A template: "AI answers vary from run to run, so we run each prompt several times and report a change only when it holds across runs. Anything else is reported as stable."

Local and e-commerce clients change the prompt set. For a local business, prompts include the city or area and are run from that location, because answers depend on where the user is. For e-commerce, add product-comparison prompts and check prices and availability for accuracy.

Where to start this week

Lock the prompt set, a couple of dozen questions the client's buyers actually ask, and the competitor set.
Choose the engines: two or three to start, where the client's buyers search.
Record the sampling conditions: logged out or API, country, language.
Run the baseline several times on different days to find the noise band.
Date and archive every run, with prompts and full answers, so later reports compare like for like.
Set up the AI channel group and the "How did you hear about us?" field.
Fill in the first one-page summary from the baseline.

To get a first baseline for these metrics, run a free Namedrop scan of your brand across ChatGPT, Perplexity, Gemini and Claude.