How to Measure GEO Success in AI Search: Mention Rate, Citation Share, Sentiment

Quick answer

Under the method in this guide, Canlah AI measures GEO success with five KPIs, each reported with its denominator beside it: mention rate (answers naming the brand divided by valid answers), citation share (answers linking the brand’s own domain divided by valid answers), recommendation position (answers placing the brand first divided by answers naming the brand), sentiment (favourable minus negative answers divided by answers naming the brand, on a stated scale) and accuracy (facts stated correctly divided by facts checked). A percentage without its divisor is not a measurement.

Canlah AI reports these five on a locked pool of buyer intents (4, 8 or 16 by tier, on the AI platforms in that tier), three runs per query per engine, re-run monthly, with branded and non-branded questions in separate denominators and the raw prompt, answer, citation list and timestamp kept as an archive the client owns. In its own self-audit of September 7, 2026, the same 236 runs produced a 100.0% branded mention rate and a 4.1% non-branded mention rate, and one set of 11 Google AI Overviews mentions produced 15.7%, 16.9% or 17.7% depending on the divisor.

Treat every figure as a range, not a rank. Answers vary between runs, engines and dates, so a single number cannot prove share of voice, causation or a future recommendation. Freeze the query pool, the peer set and the counting rule before the first run, because changing any of the three mid-quarter destroys the comparison.

Disclosure: Canlah AI publishes this guide and operates the measurement service described below. Metric definitions for Ahrefs Brand Radar, Profound, Peec AI, Semrush, AthenaHQ, Otterly.ai and Evertune come from their public product pages, reviewed September 23, 2026, and were not verified through paid accounts. The Canlah AI figures come from an anonymised self-audit of canlah.ai run on September 7, 2026, whose raw archive holds 317 evidence files.

What do GEO metrics actually measure?

GEO metrics measure five separate events inside a generated answer, plus one coverage condition across engines, and a brand can pass one while failing another. A generated answer changes by question, engine, market, language, account state and time, so every KPI below means something only once those conditions are recorded.

• Mention rate: Whether the answer names the brand in its own text, independent of any link.
• Citation share: Whether the answer attaches the brand’s own domain as a source, independent of any mention.
• Recommendation position: Whether the brand is the first pick, one of several options or a passing reference.
• Sentiment: Whether the description is favourable, neutral or negative, on a scale the report defines.
• Accuracy: Whether the category, market, price and capability stated in the answer are correct.
• Coverage: Whether the result holds across ChatGPT, Gemini, Google AI Overviews and Google AI Mode, or on one engine only.

Mention without citation means the engine knows the brand but sends no traffic. Citation without mention means the page entered the evidence pool while a competitor took the recommendation. Collapsing the two into one AI visibility score hides the difference that decides what to fix.

The five GEO KPIs and their denominators

Each of the five GEO KPIs is a fraction, and the divisor is the part vendors most often leave unprinted. The table states the divisor Canlah AI uses and the failure each one prevents.

The five AI search visibility metrics Canlah AI reports, with the denominator used for each, as published September 23, 2026.

KPI	Numerator	Denominator	Question it answers	What breaks without the stated divisor
Mention rate	Answers naming the brand in answer text	Valid answers returned, per engine, per question class	Does the engine know the brand exists	Scheduled runs counted as answers deflate the rate
Citation share	Answers attaching the brand’s own domain as a source	Valid answers returned, per engine	Does the brand’s page enter the evidence pool	Redirect URLs counted as no citation report 0.0%
Recommendation position	Answers placing the brand first	Answers that name the brand at all	When named, is the brand the pick	Dividing by all answers merges presence with ranking
Sentiment	Answers coded favourable, minus those coded negative	Answers that name the brand at all	How is the brand described	Answers that omit the brand score as neutral
Accuracy	Facts stated correctly about the brand	Facts checked in the same answers	Is the description right	Unchecked claims counted as correct

Source: the Canlah AI counting rule, published with this guide on September 23, 2026; it describes Canlah AI’s own arithmetic, not independently tested vendor arithmetic. “Valid answers” means runs that returned a complete answer a buyer could have seen, which for Google AI Overviews means a rendered overview panel, and every excluded run is listed in the report with its failure reason.

Why the denominator decides the number

The same 11 Google AI Overviews mentions give Canlah AI a mention rate of 15.7%, 16.9% or 17.7% in its September 7, 2026 self-audit, depending only on the denominator. Canlah AI scheduled 70 Google AI Overviews runs in the Singapore market. Five failed in transport and returned nothing. Three returned a results page but rendered no overview panel, and 62 produced an overview, 11 of which named the brand.

That gives 11 of 70 scheduled runs, or 15.7%. It also gives 11 of 65 runs that returned any response, or 16.9%. It also gives 11 of 62 runs where an overview appeared, or 17.7%. All three are defensible and none is wrong, so a report that prints only the percentage lets the reader assume the most flattering one.

The rule Canlah AI applies is the third, which is the valid-answer rule in the KPI table: the denominator is the set of answers a buyer could have seen. Transport failures and missing panels are reported as separate counts on the same page, because a panel that does not render is an engine behaviour worth knowing rather than an error to hide inside an average.

Count branded and non-branded questions in separate denominators

Canlah AI’s branded mention rate was 100.0% and its non-branded mention rate was 4.1% in the same September 7, 2026 archive, a 24-fold gap that one shared denominator would blur into a blended 21.2%. The archive holds 236 probes: 117 on OpenAI and 119 on Gemini, each question repeated three times.

On the 42 branded runs, where the question already named the company, the brand appeared in 42 answers, or 100.0%. On the 194 non-branded runs, it appeared in eight, or 4.1%. The blended figure across all 236 runs is 50 of 236, or 21.2%, and no buyer has ever experienced it.

The same split held on Google AI Overviews: all 11 overview mentions sat on branded questions, and on the 51 non-branded runs that returned an overview the brand was named zero times. Branded mention rate measures whether the engine holds a correct record of a company people already know. Non-branded mention rate measures whether the engine will introduce it to somebody who does not. Only the second is a growth metric.

How to count an AI citation when the engine hides the URL

The same 119 Gemini runs return a 0.0% or a 14.3% own-domain citation rate depending on which field the counter reads. Gemini returns grounding sources behind a vertexaisearch.cloud.google.com redirect, so the publisher domain never appears in the URL string. A counter matching on URLs found canlah.ai in zero of 119 runs. A counter reading the grounding title field found it in 17, all on branded questions.

OpenAI returns resolvable source URLs, so both rules agree there: the URL rule and the title rule return the same count, and the brand’s own domain appeared far more often on branded runs than on non-branded ones.

This is an instrumentation failure inside Canlah AI’s own pipeline, not a client result, and it matters because it is invisible: the URL rule produces a clean, plausible, well-formatted zero. Before any citation-share figure is reported, run one known-good query per engine and record which field the counter read.

A second structural zero comes from the table shape. In the Canlah AI archive, the 797 third-party citations across 304 domains form a peer list that excludes the own domain by construction, so reading it as total citation share would report the brand at 0.0% whatever it published.

How to score AI sentiment without borrowing somebody else’s scale

Sentiment is the one GEO KPI that vendors publish on three incompatible scales: Peec AI uses a 0 to 100 score, Profound uses three buckets and Semrush uses a driver-based Perception report. Peec AI shows the score beside visibility and position, with example brand scores from 33 to 89 on its public page. Profound labels its buckets GOOD, NEUTRAL and BAD. Semrush routes sentiment through a Perception report with a Key Sentiment Drivers section rather than a single index, and none of these numbers converts into another.

Three rules keep a sentiment figure usable. Score only the answers that name the brand, because an answer that omits it is a mention-rate event and scoring it as neutral drags every period toward 50. Code the sentence about the brand, not the tone of the whole answer. Keep the coded excerpt so a second reader can re-score it.

How named AI visibility tools label each metric

Eight AI visibility platforms, Canlah AI included, publish at least four different labels for the mention metric alone: mention rate or mentions, visibility score, visibility percentage and brand appearance. The table records each vendor’s published label, not tested behaviour, and the last column names the point to check before comparing figures.

How eight AI visibility platforms name their core metrics, public product pages reviewed September 23, 2026.

Tool	Mention metric	Citation metric	Sentiment scale published	Denominator published	What to verify before comparing
Canlah AI	Mention rate, branded and non-branded apart	Own-domain citation rate per engine, field named	Favourable, neutral or negative, over answers naming the brand	Printed beside every figure in the report	Whether the archive delivered matches the printed counts
Ahrefs Brand Radar	Mentions, how often the brand name appears in AI answer text	Citations, how often AI links to your pages or domains as a source	Not found on reviewed page	AI Share of Voice, over AI responses in your topic set	Which topic set defines the response pool
Profound	Visibility score across tracked prompts	Citation sources and their authority	GOOD, NEUTRAL and BAD buckets	Share of voice stated, divisor not found on reviewed page	The divisor behind share of voice
Peec AI	Visibility percentage per engine	Source usage at domain or URL level, with citation frequency	0 to 100 score	Visibility over tracked prompts	How the 0 to 100 sentiment score is computed
Semrush	Brand mentions in the AI Visibility Toolkit	Cited sources feeding the answer	Perception report with Key Sentiment Drivers	AI share of voice as brand mentions over all brands’ mentions	Which brands enter the all-brands denominator
AthenaHQ	Brand mention frequency tracking	Citation rate	Sentiment scores, scale not found on reviewed page	1 credit equals 1 AI response, so depth is buyer-set	How many responses per prompt the plan’s credits buy
Otterly.ai	Brand mentions across six AI platforms	Citation tracking with link citation share	Sentiment around the mention, scale not found on reviewed page	Prompts per tier, 15 to 400	Runs per prompt at the chosen tier
Evertune	Brand appearance across models	Not found on reviewed page	Not found on reviewed page	Up to 100 samples per prompt per model, across 11 models	Where citation data sits in the product

“Not found on reviewed page” means the vendor’s public materials did not name that capability when reviewed on September 23, 2026. It is not proof that the capability is absent.

Ahrefs Brand Radar publishes the cleanest separation on this list, defining mentions, citations and share of voice as three distinct metrics rather than one composite. Peec AI reports source usage at both domain and URL level, which shows the exact page an engine read. Semrush prints its share-of-voice formula as brand mentions over all brands’ mentions, so the divisor can be checked.

Profound names three sentiment buckets that a second reader can apply without a model. AthenaHQ fixes its unit at one credit per AI response, which makes sampling depth visible in the bill, and Otterly.ai publishes its prompt allowance per tier, from 15 to 400. Evertune publishes the deepest sampling figure reviewed, up to 100 samples per prompt per model across 11 models. Those published definitions, not the size of the index behind them, are what an internal reporting template should copy.

How to build the measurement record, step by step

A GEO measurement record is six decisions taken once, then repeated unchanged for at least one quarter.

Step 1: Fix the query pool and the question classes

Lock 20 to 30 buyer queries and tag each one branded or non-branded before the first run. Canlah AI’s own audit drew 39 probed queries from a universe of 191 candidates, pruning 34 and quarantining four, with every pruning decision recorded rather than deleted.

Step 2: Fix the repeat count per engine

Set at least three runs per query per engine and hold it constant. Repeats turn a single draw into a range, and a KPI built on one run per query carries no dispersion at all.

Step 3: Write the denominator rule before the first run

State for each KPI whether the divisor is scheduled runs, valid answers or answers naming the brand, and decide in advance how a broken run is treated. In the September 7, 2026 archive, 14 of 117 OpenAI runs returned truncated and one failed outright.

Step 4: Name the field the counter reads

For every engine, record which field supplies the mention string and which supplies the citation domain, then verify both against a query whose answer is already known.

Step 5: Log nine fields on every run

Store prompt, engine, model version, market, language, run index, timestamp, full answer text and the citation list. The archive, not the dashboard, is what lets a third party recompute the number.

Step 6: Report ranges and exclusions on the same page

Publish the high and low of each KPI across repeats, and list every excluded run with its failure reason beside the totals.

How to interpret five common GEO measurement results

Five GEO measurement results come up often enough to deserve a fixed reading, and each one points to a different fix, from third-party sources to the counting rule itself.

Branded mention rate high, non-branded near zero: The engines hold a correct record of the company but do not introduce it. This is the most common opening state, and it points at third-party sources, not at the brand’s own site copy.

Mention rate flat, citation share rising: The brand’s pages are entering the evidence pool before the answers name it. Keep the panel unchanged for another cycle, because this is usually the first movement in a working programme.

Citation share reported as exactly zero: Check the counting rule before the content, because a redirecting engine or a peer-only citation table produces a structural zero that publishing cannot move.

Sentiment moving while mention rate is small: Treat the swing as noise, because a score computed over four or five answers moves on a single rephrased sentence.

One engine strong, the others absent: Report engines separately and resist averaging them, because a blended score conceals which surface is working.

What GEO metrics cannot prove

GEO metrics cannot prove five things: a stable share of voice, causation, future recommendations, comparability across vendors or revenue.

• They cannot prove a stable share of voice, because the same question asked three times often returns three shortlists.
• They cannot prove that one content change caused an answer to change. Before-and-after movement is association, not causation.
• They cannot guarantee future mentions, citations or recommendations after a model or index update.
• They cannot be compared across vendors, because two platforms reporting 18% may be dividing by different things.
• They cannot replace revenue attribution. A mention rate is a visibility measure, not a pipeline measure.

When to move from a spreadsheet to a measured programme

A spreadsheet is sufficient while one person runs one panel on two engines and nobody else reads the output. Canlah AI’s protocol runs by hand: freeze the pool, repeat three times, log nine fields, print the denominator.

Three triggers justify the upgrade: the number of engines, languages, competitors or stakeholders grows until the counting rule drifts unnoticed; an inaccurate answer creates commercial or regulatory risk; a board asks for a figure an outsider can recompute. Canlah AI is less useful to a team whose only goal is one monthly mention count on one engine, which Otterly.ai’s Lite plan, listed at US$29 a month for 15 prompts when reviewed on September 23, 2026, covers at a fraction of the cost.

Whatever route you take, verify the number before you trust it. Ask for the divisor beside every KPI, recompute one figure from the raw archive and re-run three of the queries yourself on the same engines, because a percentage you cannot recompute is a chart, not a measurement.

Frequently asked questions

How do you measure GEO success?

Measure GEO success with five KPIs: mention rate, citation share, recommendation position, sentiment and accuracy, each printed with its denominator and engine on the same line. Mention rate and citation share are independent of each other, because an answer can name a brand without linking it and can link it without naming it. Branded and non-branded questions belong in separate denominators, because only the non-branded rate measures growth.

How do you calculate AI mention rate?

Divide the valid answers that name the brand in answer text by the valid answers returned for that question class on that engine, counting branded and non-branded questions separately. In the Canlah AI self-audit of September 7, 2026, the branded rate was 100.0% over 42 runs and the non-branded rate was 4.1% over 194 runs.

Is citation share the same as AI share of voice?

No. Citation share divides the brand’s own-domain citations by the answers returned, so its divisor is answers. Share of voice divides the brand’s appearances by every brand’s appearances, so its divisor is brands. A brand can hold a steady citation share while its share of voice falls, because a new competitor entered the category.

How many runs does a GEO metric need before it means anything?

A GEO metric needs at least three runs per query per engine and enough queries that the denominator is not a single digit. Evertune’s public page describes sampling a prompt up to 100 times per model, the upper end of the published range. The practical test is whether the high and low of the range are close enough to support the decision in front of you.

How do you measure AI sentiment?

Measure AI sentiment by scoring only the answers that name the brand, coding the sentence about the brand rather than the tone of the whole answer and stating the scale. Vendor scales do not convert: Peec AI publishes a 0 to 100 score and Profound publishes GOOD, NEUTRAL and BAD buckets. Keep the coded excerpt so a second reader can re-score it.

Is Canlah AI a GEO agency or a measurement tool?

Canlah AI is a Singapore-based SEO + GEO agency that measures and improves brand visibility inside ChatGPT, Gemini, Google AI Overviews and Google AI Mode, reporting AI visibility as re-verifiable ranges with timestamped evidence. Measurement arrives inside a managed engagement; after three months of the GEO service, clients can run the work themselves with Canlah All-Stars. Teams wanting only a monitoring login from the first day are better served by a platform such as Peec AI, Profound or Otterly.ai.

Method and sources

The metric definitions in the comparison table were taken from each vendor’s public product pages, reviewed September 23, 2026. No paid accounts or private demonstrations were used, and no vendor arithmetic was independently tested. The Canlah AI figures come from one anonymised self-audit of canlah.ai run on September 7, 2026 in the Singapore market: 39 probed queries across six funnel layers, three repeats per query, producing 117 OpenAI runs, 119 Gemini runs and 70 Google AI Overviews runs inside a 317-file archive. One date, one market and one brand do not establish a stable visibility rate.

• Ahrefs Brand Radar. Published definitions of mentions, citations and AI Share of Voice.
• Profound Answer Engine Insights. Visibility score, share of voice, sentiment buckets and citation authority.
• Peec AI. Visibility percentage, 0 to 100 sentiment score, position and source usage at domain or URL level.
• Semrush, How to measure AI share of voice. Share-of-voice definition and the Perception report for sentiment.
• AthenaHQ. Share of voice, citation rate and the one-credit-per-AI-response unit.
• Otterly.ai. Brand mentions, citation tracking and the US$29 Lite plan with 15 prompts.
• Evertune. Sampling depth of up to 100 samples per prompt per model across 11 models.
• Canlah AI GEO services. The locked query pool, repeat count and evidence-archive protocol described above.
• Canlah AI self-audit archive, anonymised, September 7, 2026. All first-party run counts, mention rates and citation counts in this guide.

See which of your AI visibility numbers survive a printed denominator. Run the free AI visibility audit →

Related articles

• What Are the Best AI Share of Voice Tools?
• The AI Visibility Market Has a Measurement Problem
• How to Choose an AI Visibility Tool
• We Ran Our Own GEO Audit