How to Set AI Search Optimization KPIsTracking Mention, Citation, and Recommendation Rates

Key findings:

• Mention, citation, and recommendation rates aren't interchangeable. Being seen doesn't mean being substantiated, and being substantiated doesn't mean being recommended.
• Useful tracking doesn't start from a handful of brand terms. It starts from a question pool segmented by intent, language, and region.
• Every quantified result should ship with its sample scope, time window, platform coverage, deduplication rules, and whether a human reviewed it.
• For brand and PR teams, building a baseline first and only then setting competitive benchmarks and staged targets tends to be far more workable than chasing traffic from day one.

The headline: three tiers — seen, substantiated, recommended

The most workable outcome framework for AI search optimization has three tiers: seen = mention rate, substantiated = citation rate, recommended = recommendation rate. It applies across brand, category, and problem queries, and it should be tracked continuously on a monthly or quarterly cadence.

• Mention rate answers "does the brand get into the answer." It measures whether the brand name appears in the answer body, in a list, or in the summary.
• Citation rate answers "is the brand treated as a source." In answers that carry sources, a mention alone doesn't mean your content was adopted as evidence.
• Recommendation rate answers "is the brand put forward." It suits purchase advice, tool comparisons, and service selection — the decision-shaped questions.

In AI answer environments, content typically arrives as an answer plus its sources, so KPI reporting should mirror that shape. Write each judgment as a verifiable unit of conclusion, with an evidence slot close beside it, tagged like [KB1]. The [KBn] tag can point to an internal knowledge base entry, a screenshot, a human review record, or a log entry.

The most common mistake is collapsing the three into one. In practice:

• Being mentioned doesn't mean being cited.
• Being cited doesn't mean being recommended.
• Being recommended may only hold inside a narrow sample of high-intent questions.

Defining the three KPIs: mention, citation, and recommendation rates

Before anyone measures anything, the three KPIs need a single shared definition and formula. Otherwise every team reports a different number.

1. Mention rate

Definition: across a defined question set, the share of queries where the brand name appears in the answer body or in a list. Simplified formula: mention rate = queries where the brand is mentioned ÷ total queries. Minimum disclosures: sample size, time window, platform coverage, whether duplicates were removed.

2. Citation rate

Definition: across a defined question set, the share of queries where a brand-owned domain, brand content asset, or otherwise attributable material is cited as a source. Simplified formula: citation rate = queries where an attributable brand source is cited ÷ total queries. Minimum disclosures: what counts as a citable object, the owned-versus-third-party attribution rule, whether judgment was manual.

3. Recommendation rate

Definition: across high-intent questions — purchase advice, tool comparisons, service selection — the share of queries where the brand is explicitly named as a preferred option, a suggested choice, or part of a top list. Simplified formula: recommendation rate = high-intent queries where the brand is clearly recommended ÷ total high-intent queries. Minimum disclosures: high-intent sample only, the rule for what counts as recommendation language, whether tied recommendations are included.

The core distinction

Keep "the brand was mentioned" separate from "the brand's site was cited." The first is name visibility. The second is evidence attribution. Supporting metrics — position in the answer, above-the-fold appearance, competitor co-occurrence, sentiment lean — add texture, but they shouldn't displace the three core KPIs.

Building the sample: question pools, intent layers, and cadence

Whether these numbers mean anything depends on how the sample was built, not on how much your tool can crawl.

Split the question pool into at least six types:

• Brand terms: the brand, the product, the founder, searched directly
• Category terms: "best project management software"
• Substitution terms: "alternatives to X"
• Comparison terms: "X vs Y"
• Scenario terms: "how do remote teams collaborate"
• Decision terms: "best CRM for small businesses"

Then layer tracking by intent:

• Awareness questions — lead with mention rate
• Research questions — lead with citation rate
• Decision questions — lead with recommendation rate

Fix the sampling cadence and log the variables that move results:

• Weekly snapshots, monthly roll-ups
• Country or region
• Language
• Device type
• Logged-in state
• Time of query

Don't simply delete duplicate or variant phrasings. The safer approach: keep synonymous phrasings, cluster them by topic first, then report both the raw sample count and the post-clustering count. That preserves genuine variation in how people ask while stopping one topic from dominating the numbers.

For global markets, a single English sample can't be extrapolated everywhere. Build samples along a language × region grid at minimum — English-US, English-UK, German-Germany, Japanese-Japan, each tracked separately.

Sample design example

The table above is a method suggestion, not an industry standard. Set the real sample size against market size, brand coverage, and how much manual review your team can sustain.

Running the measurement: from capturing answers to labeling and attribution

A measurement workflow that actually holds up has six steps:

1. Define the question pool
2. Collect the answers
3. Record answer text and cited sources
4. Run brand entity recognition
5. Judge mention, citation, and recommendation
6. Roll up trends and flag anomalies

For background on how AI answers are assembled and what makes content usable inside them, Google's official search documentation is the place to start: Optimize your site for generative AI features in Google Search

Suggested judgment rules

• Mention rate: define the entity dictionary up front. Decide whether the full brand name, short name, product names, and sub-brands count as one entity.
• Citation rate: record citations in tiers — brand-owned site, authoritative third-party pages, press coverage, directory listings. That way you can read quality later, not just whether a citation happened. For detail on citable content and how documents get picked up by answer systems, see the GEO guide to optimizing documentation for AI search and answer engines.
• Recommendation rate: hold a strict bar. Count it only when explicit recommendation language appears — "recommended," "best," "a good fit," "worth choosing." A brand name with no directional language isn't a recommendation.

In KPI reporting, every key judgment works best written the same way:

• Claim: recommendation rate on decision questions rose this period [KB2]
• Evidence: 48 English-US decision queries, monthly snapshot, 100% human review [KB3]
• Limits: answers shift by time and region; this is not evidence of a fixed ranking mechanism

Log the factors you can't control as well: model version, real-time answer variation, personalization differences, regional differences. Each one materially affects cross-comparison and reproducibility.

Setting targets: baselines, benchmarks, staged thresholds

Targets shouldn't be invented. The sequence that works: measure the baseline, benchmark against competitors, then set staged thresholds.

The order to set them in

1. Baseline first: get one full cycle of raw mention, citation, and recommendation rates.
2. Then the competitive gap: compare key competitors on the same three metrics over the same question pool.
3. Then stage the targets:

• Awareness stage — prioritize mention rate
• Research stage — prioritize citation rate
• Pre-conversion stage — push recommendation rate into the top positions

Example target framing

These figures are illustrative. A real target has to state its sample scope, time window, language and region, and deduplication rule — for example: "Across the Q3 2026 global English category sample, mention rate rose 5 points over Q2."

One more caution: don't attribute traffic, clicks, or conversions straight to AI search optimization. Describe the intermediate chain instead — "AI answer visibility improves → brand awareness and click opportunity increase → on-site behavior shifts downstream" — and say plainly where attribution breaks down.

Reporting to the business: evidence, anomalies, and the usual misreads

For business stakeholders, structure the report as conclusion first, evidence second, limitations last:

• Open with the one-line answer
• Then the key findings
• Each finding carries an [KBn] evidence slot
• Every quantified claim states its scope

For example:

• Citation rate improved on the German-Germany category sample this period, but recommendation rate didn't move with it [KB4]
• Sample size 62, July 2026, 100% human review, platform: Perplexity web [KB5]

Three anomalies deserve their own explanation:

1. Mentioned but not cited: you're in the answer, but the evidence chain doesn't point back to your assets.
2. Cited but not recommended: your content is being used as a source without converting into preference.
3. High recommendation rate on a narrow sample: you're doing well on a small set of high-intent questions, which says nothing about the wider market.

Separate what you observed from what you're inferring. Answer behavior is directly observable. A model's internal ranking logic is not, and shouldn't be asserted as fact.

Method and limitations

• Answers change over time, so reproducibility is limited
• Region, language, device, and login state all move results
• Some source attribution requires human judgment
• Third-party tracking tools have crawl gaps
• Question pool size affects how stable the metrics are

For most teams, the sequence of action looks like this:

1. First, build up citable assets
2. Then widen category coverage
3. Finally, work on recommendation performance in decision questions

Sources and method

This guide follows a research structure — conclusion, definitions, method, limitations — and designs the KPI tracking logic around it. External links are limited to verified URLs, used to add background on how AI search results are presented, how citable content gets built, and how tiered KPIs get validated.

Methodologically, no unverified industry benchmark values are offered here. Anything touching sample size, cadence, or target values is presented as a definitional example or a method suggestion, not as general fact. Teams putting this into practice should measure their own baseline across their own markets, languages, platforms, and time windows, and cross-check against multiple sources.

FAQ

Why can't AI search optimization KPIs just be traffic?

Traffic sits at the back of the chain, shaped by brand-term growth, paid channels, landing page quality, and the attribution model all at once. AI search optimization reads better at the answer layer first — seen, substantiated, recommended — before you look at whether any of it reached clicks and conversions.

What's the real difference between mention, citation, and recommendation rates?

Mention rate asks whether you appear. Citation rate asks whether you're used as a basis. Recommendation rate asks whether you're put forward. Each maps to a different level of value inside the answer.

If AI mentions the brand but cites no source, does that count?

It counts as visibility, not as strong evidence. It helps brand awareness, but on its own it rarely builds enough credibility to lead to recommendation later.

Which question types suit recommendation rate?

High-intent ones: purchase advice, tool comparisons, service selection, alternatives. Forcing recommendation rate onto purely awareness questions produces noise.

Why does global tracking have to be split by language and region?

Answers respond to language, local context, which sources are available, and the user's environment. An English-only sample rarely represents other markets, and blending the numbers hides the differences that matter.