Recommendation Rate: The Metric That Matters

Recommendation rate is the share of relevant AI answers that suggest your brand as an option for the user's need, not just name it. Calculate it as answers labelled "recommended" ÷ all answers for the prompt group, on unbranded discovery and comparison prompts, per engine, with the sample size.

On this page

• How do mention, citation and recommendation differ?
• How do you define a recommendation?
• How do you calculate it?
• Which supporting dimensions explain the number?
• Who should label answers?
• From MentionX data: where recommendations can happen
• How do you improve recommendation rate?
• Worked example: Ledgerline fixes a stale caveat
• How do you score position inside a recommendation?
• How should you report recommendation rate?
• What should you do when a competitor is recommended instead?
• How does it relate to revenue?
• How do you keep labels consistent?
• Common mistakes
• Checklist
• Where to go next
• How do engines differ?
• Common questions
• Sources

Engine documentation checked on 30 October 2026. "Our view" marks MentionX opinion. Ledgerline is fictional.

How do mention, citation and recommendation differ?

The first example mentions Acme without recommending it. A citation can also occur without a recommendation, for instance when your blog post is used as a source for a general fact.

How do you define a recommendation?

Write a short labelling guide. A sample definition (illustrative):

A recommendation occurs when the answer:

• includes the brand in a list of suggested options for the user's need, or
• explicitly advises considering, trying or choosing the brand, or
• names the brand as best for a stated segment ("best for small teams").

It's not a recommendation when the brand is:

• mentioned only as a comparison point or something to avoid,
• named only in the sources, or
• described with a caveat that advises against it for this need.

Give each answer one label: recommended, mentioned-not-recommended or absent. Edge cases go in a "notes" column and get reviewed.

How do you calculate it?

Recommendation rate = answers labelled recommended ÷ total answers for the prompt group.

Illustrative: 100 discovery prompts in ChatGPT; your brand is recommended in 22, mentioned without recommendation in 9 and absent in 69. Recommendation rate = 22%; mention rate = 31%. Report the range too: at n=100 a 22% rate has a 95% Wilson interval of roughly 15%–31% (why ranges matter).

Which supporting dimensions explain the number?

Reasons and segments are often the most useful: if every answer recommends you "for large enterprises" while your target is SMBs, you have a positioning problem to fix in content and third-party coverage.

Who should label answers?

Manual labelling is accurate but slow; automated classification is fast but needs checking. A practical approach: automate first-pass labels, then have a person review a random sample each period and all edge cases. Record agreement rates; if humans disagree with the automated labels often, refine the definition.

From MentionX data: where recommendations can happen

Method. Read-only aggregate queries over 2,154 completed MentionX runs from 22 August to 29 October 2026 (UTC), across 18 anonymised projects, mostly B2B software, and six engines called through APIs. We show mention rates and share of voice by prompt intent, because a brand can only be recommended in answers that mention it.

Where the tracked brand appeared in a ranked list (n=102 placements): first 18%, second or third 14%, fourth or fifth 13%, sixth or lower 56%.

What it means. Comparison prompts almost always mention the brand, because the prompt names it. Recommendation and best-of prompts, where buyers ask what to choose, mention it in under a third of answers. So recommendation rate on those prompts can't exceed roughly 30% for a typical tracked brand here, and in practice it sits lower once caveats and "avoid" mentions are removed. Most list placements also sit low, which is weak even when the answer technically counts as a recommendation.

Limits. A small, non-random sample of brands that chose to track themselves. API answers can differ from consumer apps. We didn't publish a recommendation rate here because labelling rules differ between projects; the mention figures are the consistent upper bound.

Our view: track recommendation rate only on prompts where the buyer is choosing. On comparison and branded prompts it measures little except whether the engine read the question.

How do you improve recommendation rate?

1. Read the reasons. Find which criteria answers use (price, integrations, support) and check your content states your position on each clearly.
2. Close gaps in third-party evidence. Assistants often reflect reviews, comparisons and community discussion. Make sure accurate information about you exists there.
3. Publish honest comparison and "best for" content. Explain who you're right for, and who you're not.
4. Fix inaccuracies at the source. Outdated pricing or features on your site or review profiles can cause caveats.
5. Measure per segment. Improvements often show up first for the segment you target.

Worked example: Ledgerline fixes a stale caveat

Ledgerline is a fictional accounts-payable tool. Figures are illustrative. Ledgerline tracks 80 unbranded discovery prompts in ChatGPT and Perplexity. In ChatGPT it's mentioned in 40% of answers but recommended in only 15%. Reading the reasons, the team sees a pattern: answers name it, then add "but it lacks multi-entity approvals". That was true two years ago; the feature shipped last spring, but the pricing page, two review profiles and an old comparison article still say otherwise.

The team updates its feature pages with a clear, dated statement, asks review sites to refresh the listing, rewrites the comparison article and publishes a short changelog entry. Three months later, the caveat appears less often in the labelled sample. The team reports the change with ranges, noting that the intervals overlap only slightly and that it'll confirm with another quarter of data before claiming success.

How do you score position inside a recommendation?

A first-place recommendation and a ninth-place one both count as "recommended" under a simple definition. If position matters in your category, add a weighted view next to the plain rate:

Weighted recommendation rate = sum of weights for recommended answers ÷ total answers. Publish the weights and keep them fixed. Report both figures: the plain rate is easy to explain, and the weighted one shows whether you're moving up lists, which matters given how many placements sit far down.

How should you report recommendation rate?

Keep the report short and consistent each month:

1. Recommendation rate per engine, with n and a range.
2. Change from last period, marked "significant" only if tested.
3. Top three reasons you're recommended, and the top caveat.
4. Competitors most often recommended instead.
5. One action for next month.

Leaders tend to remember the reasons and caveats more than the percentage, and those are what teams can act on.

What should you do when a competitor is recommended instead?

Read ten answers where a competitor is recommended and you're not, or you're mentioned without a recommendation. Note the reason the answer gives for the competitor and the sources it cites. Three patterns come up most:

1. A clear segment claim. "Best for mid-market teams" appears on their site and on review pages. If you serve that segment too, say so plainly and with evidence.
2. Third-party presence. They appear on review and comparison pages the engines cite, and you don't. Get listed accurately on those pages.
3. A feature named in the question. If buyers ask for a feature you have, make sure it has its own clearly titled section or page.

Pick one pattern per quarter and track its prompts separately, so you can see whether the fix moved anything.

How does it relate to revenue?

Recommendation rate is a leading indicator, not revenue. Connect it to outcomes with AI referral sessions in analytics, self-reported attribution on forms and sales conversations. If all three rise together over a few quarters, the case that AI recommendations are driving demand gets much stronger. If recommendation rate rises but nothing downstream moves, check whether you're being recommended for the right segment.

How do you keep labels consistent?

Labelling is where recommendation rate goes wrong, because two reasonable people can read the same answer differently. A few habits keep it steady:

• Use a reference set. Keep 30 labelled answers covering clear cases and edge cases. New reviewers label them first, and you compare.
• Measure agreement. Each month, have two people label the same 20 answers. If they agree on fewer than about nine in ten, tighten the definition before trusting the trend.
• Write rules for common edge cases. For example: "Recommended only for a segment you don't target" counts as recommended, with the segment noted. "Listed under 'other options'" counts as recommended, with position recorded. "Good, but expensive" counts as recommended with a negative caveat.
• Label blind to the month. Mix answers from this month and last month in one batch so reviewers can't lean towards the result they expect.
• Restate history when the definition changes. Re-label the previous three months under the new rule, or start a new trend line.

Our view: agreement between reviewers is worth reporting next to the rate. It tells readers how much of a small change could be labelling noise.

Common mistakes

• Counting every mention as a recommendation.
• Measuring only branded prompts, where you're almost always recommended.
• Changing the definition mid-year without restating history.
• Celebrating one month's jump without checking the interval.

Checklist

• Written definition with examples and exclusions
• Three-way labels: recommended / mentioned / absent
• Calculated per engine and prompt group with n
• Position, segment, reason, sentiment, competitors recorded
• Human review of a sample each period

Where to go next

• AI visibility KPIs for marketing teams
• Building a prompt set
• Statistical confidence for AI visibility tracking

How do engines differ?

Rates differ a lot by engine, so never average them without showing each one. In the same MentionX data, tracked brands were mentioned in 47.3% of Gemini answers, 42.9% of OpenAI, 31.5% of Perplexity, 31.3% of Google AI Mode, 26.0% of Copilot and 24.0% of DeepSeek answers (Perplexity, Copilot and DeepSeek from a single day, 26 October). Engines also differ in how they present options: some give short ranked lists, others longer prose with several "it depends" branches. Write labelling examples for each engine's format so reviewers apply the same rule.

Common questions

How is recommendation rate different from mention rate? A mention counts any appearance. A recommendation needs the answer to suggest your brand as an option or choice for the user.

What counts as a recommendation? Whatever your written definition says, for example inclusion in a list of suggested options or a direct "consider X". Keep it stable for the year.

Can a brand have a high mention rate but low recommendation rate? Yes, for example when it's mentioned as an expensive alternative or described with caveats.

Can a model label the answers? Yes, as a first pass. Check a random sample by hand each month and record how often people agree with the labels.

Sources

• NIST/SEMATECH, Confidence intervals for proportions
• Google Search Central, AI features and your website, on how AI features present and link sources
• OpenAI Help, ChatGPT search