AEO KPIs that replace keyword rankings: an agency guide to reporting AI visibility to clients

KPI 1 · Visibility

Mention rate

The share of tracked prompts where the AI engine names your client's brand in its answer.

Mention rate = answers naming the brand ÷ total scored answers in the tracked set × 100

How to sample it: Run a frozen set of 30 to 100 buyer prompts at least three times per prompt per engine, weekly, and count a prompt as a mention if any run names the brand. Report the rate per engine, per month.

Realistic starting range: On non-brand buyer prompts, expect single digits at the start: across the three UK B2B sites we scan, the median non-brand mention rate is 6%, and the spread runs from under 1% to 23%. Brand-name prompts sit near 100% and belong on a separate line.

KPI 2 · Visibility

Citation rate

The share of tracked answers that link to or source your client's domain, not just name the brand.

Citation rate = answers citing the client's domain ÷ total scored answers × 100

How to sample it: Capture the source list attached to each answer (Gemini, Perplexity and Google AI Overviews expose one; ChatGPT often does not) and match hosts against the client's domain. Same frozen prompt set as mention rate, so the two lines compare.

Realistic starting range: Always lower than mention rate, and much lower on ChatGPT: in our August 2026 scans, 46% of ChatGPT answers that named a monitored brand carried no link to its domain, against 8% on Gemini. A citation rate of half your mention rate is a normal starting point.

KPI 3 · Visibility

AI share of voice

Your client's percentage of every brand mention across the tracked answer set, competitors included.

AI share of voice = (client's mentions ÷ total brand mentions in the same tracked answers) × 100

How to sample it: Count every brand each answer names, not just the rivals the client asked you to watch: the engines routinely recommend businesses neither of you had listed, and those mentions belong in the denominator. The full method, with a worked example, is in our share of voice guide. See the full share of voice method.

Realistic starting range: Zero-sum by construction, so benchmark against the best rival in the same answer set rather than an industry average. A new programme typically starts below its shortlist rivals; the useful target is the named competitor directly above you.

KPI 4 · Visibility

Answer inclusion rate by prompt bucket

Mention rate split by prompt type: brand, soft-brand and non-brand buyer questions reported as separate lines.

Inclusion rate per bucket = answers naming the brand in that bucket ÷ scored answers in that bucket × 100

How to sample it: Tag every prompt in the set as brand (contains the client's name), soft-brand (their category plus a distinctive detail) or non-brand (pure buyer question), and never blend the buckets. The blended figure flatters, because brand prompts are easy.

Realistic starting range: Brand prompts near 100%, soft-brand somewhere in the middle, non-brand in single digits at the start. The non-brand line is the one that predicts new business, so it leads the report even when it is the ugliest number on the page.

KPI 5 · Visibility

Recommendation position

Where in the answer the client appears when the engine lists providers: first named, third, or last.

Recommendation position = median list position of the brand across answers that name it in a list

How to sample it: For every answer that returns a list of providers, record the client's position in it. Track the median rather than the mean, because one stray answer that names them tenth should not swamp a month of second places.

Realistic starting range: Being named at all comes first; position work comes second. A brand new to the answers typically enters mid-list. First-named matters because assistants often summarise their own list, and the summary keeps the first two or three names.

KPI 6 · Visibility

Sentiment and accuracy

How favourably each answer describes the client, and whether the facts in the description are right.

Sentiment split = positive, neutral and negative answers as a share of answers naming the brand; accuracy = factual errors found per month, listed

How to sample it: Read (or machine-score, then spot-check) every answer that names the client, and log the quoted sentence behind each negative call so the client can see the evidence. Log every wrong price, retired product and mixed-up namesake as a named error, with the engine and prompt.

Realistic starting range: Most B2B answers are neutral; a handful of negatives is normal and fixable. Accuracy errors are common where pricing changed or a namesake exists. Every accuracy error is a work item, which is exactly why it belongs in a KPI report.

KPI 7 · Outcome

Cross-engine coverage

How many of the six engines (ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, AI Mode) mention the client on non-brand prompts.

Cross-engine coverage = engines with a non-zero non-brand mention rate ÷ engines tracked × 100

How to sample it: Same prompt set, run per engine, scored per engine. Never average engines into one blended figure: the engines source answers differently, so a fix that moves Gemini often does nothing for ChatGPT, and the per-engine lines are what tell you where the problem lives.

Realistic starting range: Most clients start visible on one or two engines and absent on the rest, because Gemini and AI Overviews lean on the live Google index while ChatGPT leans on training data plus occasional search. Six out of six on non-brand prompts is rare and worth reporting loudly.

KPI 8 · Outcome

AI referral traffic and assisted conversions

Sessions, key events and revenue from visitors arriving off AI assistant domains, measured in the client's analytics.

AI referral share = sessions from AI assistant sources ÷ all sessions × 100, with conversions counted on the same segment

How to sample it: GA4 shipped a default AI Assistants channel in May 2026: start there, then build a custom channel group on a referrer regex for chatgpt.com, perplexity.ai, claude.ai, gemini.google.com and copilot.microsoft.com, because custom groups apply retroactively and give you history. Treat the number as a floor: assistants strip referrers, and some readers type the brand into Google instead.

Realistic starting range: Small but growing fast: ChatGPT referrals to tracked B2B sites roughly quadrupled in the year to June 2026 (Labs by Demandbase). Expect a low single-digit share of sessions that converts respectably, and chart it monthly next to organic.

KPI 9 · Outcome

Gap-to-fix closure rate

The share of identified visibility gaps that were actually fixed this month, on the client's site or off it.

Closure rate = gaps fixed and shipped this month ÷ gaps identified and accepted × 100

How to sample it: Keep a fix log: every gap the scans surface, who owns it, the date it shipped, and the re-scan result afterwards. Count a fix as closed when the change is live, and separately track whether the answers moved at the next scan.

Realistic starting range: This is the KPI that separates a reporting agency from a fixing one, and the client's clearest evidence of work done. A steady cadence of a handful of shipped fixes a month, each tied to a named gap, beats a long list of identified-but-untouched items every time.

Mention rate falls

Likely cause: A model or retrieval change on the engine's side, or a rival newly present in the sources the engine reads. Rarely something you broke this month.

First fix: Check the run log against engine release dates first, then read the citations on the changed answers. If a new source is doing the recommending, get the client onto that source.

Citation rate falls while mentions hold

Likely cause: The engine still knows the brand but has found better-structured pages to evidence with, or the client's pages have become unreadable to crawlers.

First fix: Re-check robots.txt and llms.txt allow the AI crawlers, then compare the winning cited pages against the client's: usually the rival page answers the question in its first paragraph and the client's does not.

Share of voice falls

Likely cause: A competitor shipped something the engines picked up: a roundup placement, a comparison page, a burst of reviews. Your client's mentions can hold steady while their share still falls.

First fix: Read the answers that changed and name the rival taking the share. The fix list is usually their citation list: pitch the roundup, publish the comparison, earn the reviews.

Non-brand inclusion stalls at zero

Likely cause: The engines have no page anywhere that connects the client to the buyer question, so there is nothing to retrieve and nothing to remember.

First fix: Publish a direct answer to each stalled prompt on the client's site (question as heading, answer in the first sentence), then get the client named in one third-party source the engines already cite for that question.

Recommendation position slips

Likely cause: The engines cite comparison and best-of pages in order, and the client has slipped down the underlying lists, or a rival has earned a stronger differentiator sentence.

First fix: Give the engines a reason to put the client first: a specific, dated claim on the client's own pages, repeated by a third party. Vague superlatives do not travel; numbers do.

Sentiment turns negative

Likely cause: The engine is quoting a real complaint (a review thread, a forum post) or misreading old pricing and stock as current.

First fix: Find the quoted source, fix what is fixable at the source, and publish the correction plainly on the client's site. Then re-scan until the answer changes; sentiment moves when its evidence moves.

Accuracy errors appear

Likely cause: Stale facts on pages the engines read: an old price on a forgotten landing page, a dead product in a directory listing, a namesake business being blended in.

First fix: Correct the stated fact everywhere it appears in visible body text, on the client's site first and the third-party listings second. Plain, current, visible text is what fixes stale claims.

AI referral traffic dips

Likely cause: Engine-side link behaviour changes month to month (ChatGPT's link-out rate has moved severalfold within a year), so the line moves for reasons that have nothing to do with visibility.

First fix: Check visibility KPIs before reacting: if mentions and citations held, the dip is plumbing, not presence. Note the engine-side change in the report and keep the trend line.

Closure rate falls

Likely cause: The fix list has outgrown the people fixing it, or gaps are being identified faster than anyone agreed to resolve them.

First fix: Cut the list, not the standard: agree a monthly fix budget with the client (a number of shipped fixes), rank gaps by the revenue closeness of their prompts, and carry the rest visibly rather than silently.