Measurement · By Asfand Yar Junejo · 13 min read

Ranking, retrieval, citation: why measure them separately?

Ranking, retrieval, mention, citation, recommendation and conversion are six different outcomes, and one does not guarantee the next. A page can rank without being retrieved by an AI assistant, be retrieved without being cited, and be cited without anyone clicking. Measuring them as one “visibility” number hides where the chain actually breaks.

Cover graphic: ranking, retrieval, mention, citation, recommendation and conversion as six separate AI visibility outcomes, each measured on its own
Six different outcomes, six different fixes: measure the chain, not a blended score.

What does each outcome mean?

Each one describes a different moment between a question being asked and a customer acting on the answer.

OutcomeWhat happened
RankingYour page appeared at a position in a list of results
RetrievalA system pulled your content into what it used to build an answer
MentionThe answer named your brand
CitationThe answer linked your page as a source
RecommendationThe answer suggested you as the choice
ConversionSomeone arrived and did what you wanted them to do

Which outcomes can actually be measured?

Rankings and conversions can be measured directly; mentions and citations can be observed by testing; retrieval mostly has to be inferred.

OutcomeHowCertainty
RankingSearch Console, rank trackersMeasured
RetrievalRarely exposed by platformsInferred
Mention / citationRepeated, logged prompt testsObserved
RecommendationDecision-stage prompt testsObserved
ReferralAnalytics referrer data, partlyPartly measured
ConversionGA4 or CRM eventsMeasured

How do you test AI answers properly?

Treat every prompt test as an experiment: fix the conditions, repeat it, log everything and compare like with like.

  1. Write prompt families, not single prompts: informational, comparison and decision-stage questions
  2. Record the conditions: platform, model if shown, date, region, signed in or not, search mode
  3. Separate cold and contextual tests: a fresh chat behaves differently from a long conversation
  4. Repeat each prompt several times, because answers vary run to run
  5. Log mentions, citations and recommendations as separate columns
  6. Re-run monthly and date every result, since platforms change without notice

How do you log all six outcomes in one prompt test?

Test each query twice, in a cold session and in a contextual follow-up, and record the six outcomes in separate columns instead of one combined score.

A cold session means a clean browser profile, logged out, with no earlier conversation. This is the blank logging template; it contains no results.

ColumnWhat to record
Prompt and sessionThe exact wording, copied, and whether the session was cold or a contextual follow-up
1. RankingSearch Console average position for the target URL
2. RetrievalWhether the page was fetched or a passage from it was used
3. MentionWhether the brand is named in the answer text
4. CitationWhether the answer links to the domain
5. RecommendationWhether the brand is positioned first or presented as the pick
6. ConversionThe click or inquiry, traced to its source

A downloadable Prompt Tracker for this is on the measurement method page, with the columns to record each run.

How do you separate AI citations from search clicks in GA4 and Search Console?

Use two tracks: referral sessions in GA4 for the visits an assistant sends, and branded-query trends in Search Console for the citations that send no click.

  1. AI referral sessions (GA4). Filter session source with a pattern for the assistants you track, for example chatgpt\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com. Compare conversion rate and engaged time with organic search rather than assuming either is higher. Referrer names change, so review the list regularly.
  2. Zero-click citations (Search Console plus the prompt log). A citation without a click often shows up as more exact-brand and “brand + service” queries. Compare weekly branded impressions against the dated prompt-log entries.

Before you trust either track, know why GA4 sessions and Search Console clicks never match; this guide to connecting Claude to both safely explains the differences and how to reconcile them.

Directional Branded-query growth is consistent with, but does not prove, an AI citation as the cause

What are the common measurement mistakes?

Treating one screenshot as a trend, and treating a correlation as a cause.

  • Reporting a single AI answer as proof of visibility
  • Adding mentions and citations together into one score
  • Crediting a change for a rise that also happened to competitors
  • Comparing results taken on different platforms or dates
  • Hiding tests that came back negative or inconclusive

The GEO Visibility Index in the Lab applies this routine to citation tracking.

Why does one "AI visibility" score hide the real problem?

Because each outcome has a different cause and a different fix, and a blended score cannot tell you which link in the chain failed.

Imagine a brand that is not named in an AI answer. The cause could be that the page is not crawlable, that it is crawlable but never ranks for the question, that it ranks but its passage does not answer the question cleanly, or that the passage is used but the brand name never appears in it. Those are four different jobs for four different people: a developer, an SEO, an editor and a brand manager. A single "visibility: 12%" figure points at none of them.

Separating the outcomes turns a vague complaint ("AI does not mention us") into a diagnosis ("we rank, we are cited, but the answer recommends a competitor"). That is a much shorter route to a fix, and it also stops you taking credit for a good number that has nothing to do with your work.

Confirmed Google says generative AI search is still search, so ordinary crawl, index and quality practices apply to the first links in the chain. Directional How the later links behave differs by platform and changes without notice.

What does each link in the chain depend on?

Ranking depends on classic search signals, retrieval on how an assistant picks and reads sources, mention and citation on what the passage says, recommendation on how the answer is framed, and conversion on the page the visitor lands on.

What each of the six outcomes depends on, and the first thing to check when it is missing
OutcomeMainly depends onFirst thing to check when it is missing
RankingCrawlability, indexing, relevance, quality and linksIs the URL indexed, and for which queries does Search Console show it?
RetrievalWhether a system can fetch the page and finds a passage worth usingIs the page blocked by robots.txt or login, slow, or rendered only by script?
MentionThe brand name sitting next to the answer in the source textDoes the passage name the brand, or only say "we"?
CitationA page the assistant treats as a source for that claimIs there a page that answers this exact question, with a stable URL?
RecommendationComparative evidence, reviews, and how the question is framedDo independent sources support the claim that you are the right choice?
ConversionLanding page clarity, offer fit and trustDoes the landing page answer the question the visitor arrived with?
Diagram of the six AI visibility outcomes in order: ranking, retrieval, mention, citation, recommendation and conversion, each measured and diagnosed separately
Figure 1. The six outcomes as a chain: each link is measured on its own and fixed by a different person.

What is the difference between a mention, a citation and a recommendation?

A mention names your brand, a citation links to your page as a source, and a recommendation positions you as the choice; an answer can contain any one without the others.

  • Mention without citation: the assistant knows your name from training data or other sites but does not link to you. You gain awareness, not a visit.
  • Citation without mention: your page is listed as a source, but the answer text never says your name. You may get a click without brand recall.
  • Mention with a negative frame: the answer names you as an example of what to avoid. Counting it as a win would be a reporting error, so record the sentiment in its own column.
  • Recommendation: the answer presents you as the pick for a decision-stage question such as "which provider should I choose for..."; this is the outcome closest to revenue and the hardest to earn.

This is why the log template in this post keeps three separate columns. When a client asks for "citations", ask which of the three they mean, because the work and the evidence differ for each.

Why do AI answers change between runs, and how should you handle it?

Generated answers are not fixed documents: wording, sources and brand lists can vary between runs, accounts, locations and dates, so one result is an anecdote and only repeated, dated tests show a pattern.

Treat a prompt like a survey question. Ask it several times under the same conditions, record how often each outcome appears, and report that as a rate with its date range rather than as a yes or no. When the rate moves, check whether the platform, model or search mode changed before you credit your own work.

  1. Fix the conditions you control: same wording, same session type, same region, same search mode
  2. Run each prompt more than once and keep every run, including the ones that do not mention you
  3. Report a rate and a range ("appeared in 3 of 5 runs, 1 to 8 November"), never a single screenshot
  4. Note platform changes in the log so a jump or drop can be explained
  5. Include competitors in the same prompts so you can tell a market-wide shift from a change that only affects you

Directional Run-to-run variation is widely observed by practitioners; this post gives no universal number of runs because none has been established.

How do you design prompt families that match the buying journey?

Write prompts for each stage of the decision, because informational, comparison and decision-stage questions pull different sources and test different outcomes.

Prompt families by decision stage, with the outcome each one mainly tests
StageExample prompt patternOutcome it mainly tests
Learn"What is [topic] and how does it work?"Retrieval and citation of explanatory pages
Compare"[Option A] vs [Option B] for [use case]"Mention, plus which sources frame the comparison
Shortlist"Best [service] providers for [audience]"Mention and recommendation
Decide"Is [brand] a good choice for [need]?"Recommendation and sentiment
Act"How do I contact or buy from [brand]?"Referral and conversion path

Keep five to ten prompts per stage that real customers use, taken from sales calls, support tickets and Search Console queries, not invented from your own vocabulary. If you are building briefs from customer language, turning customer interviews into an AI-assisted content brief shows how.

How do you read the results and decide what to fix first?

Find the first link in the chain that fails, fix that, and only then move to the next, because improving a later link does little while an earlier one is broken.

Common result patterns, what they usually indicate and the first fix to try
Pattern in the logLikely meaningFirst fix to try
No ranking for the target queryIndexing, relevance or quality gapCheck indexing and the page's match to the query in Search Console
Ranks, but never citedThe passage does not answer the question cleanlyRewrite the section answer-first with the subject named
Cited, but brand not namedThe source text never ties the answer to your nameName the brand in the answer sentence and on the entity page
Named, but not recommendedWeak independent corroboration or unclear fitAdd evidence, case detail and third-party references; state who it is for
Cited, but few visits convertLanding page does not match the answer's promiseAlign the landing page headline and first paragraph with the cited claim
Name is wrong or mixed with another entityEntity signals are inconsistentBuild a single entity home page and align profiles

Hypothesis The "likely meaning" column is a diagnostic starting point from practice, not a documented rule; verify each against your own data before acting.

How should you report the six outcomes to a client or a manager?

Report each outcome separately with its method and certainty label, show the trend against the prior period, and end with the single next action for the weakest link.

  • One row per outcome, each tagged measured, observed or inferred
  • Date range and sample size next to every prompt-based number
  • Competitors in the same table, so context is visible
  • Negative and inconclusive results kept in; hiding them makes the next report untrustworthy
  • One recommended action, tied to the first failing link

For the reporting mechanics in GA4 and Search Console, see AI visibility tracking: citations, referrals and qualified leads and automated weekly SEO reporting. Google also describes a Generative AI performance report in Search Console in its optimization guide, which is the first-party source to check before trusting a third-party number.

Frequently asked questions

Can a page rank without being cited by an AI assistant?

Yes. Ranking, retrieval, mention, citation, recommendation and conversion are six different outcomes. A page can rank without being retrieved, be retrieved without being cited, and be cited without anyone clicking.

Which of the six outcomes can be measured directly?

Rankings and conversions can be measured directly (Search Console, rank trackers, GA4 or CRM events). Mentions and citations can be observed by repeated, logged prompt tests. Retrieval is rarely exposed by platforms, so it is mostly inferred.

What is the difference between an AI mention and an AI citation?

A mention means the answer names your brand in its text. A citation means the answer links to your page as a source. Either can happen without the other, so log them in separate columns.

How many times should an AI prompt be tested?

Several times, because answers vary from run to run; the post gives no fixed number. Record the platform, date, region, signed-in state and search mode, and re-run monthly with dated results.

How do you find AI referral traffic in GA4?

Filter session source with a pattern for the assistants you track, for example chatgpt\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com, and review the list regularly because referrer names change. Compare conversion rate and engaged time with organic search rather than assuming either is higher.

Does a rise in branded searches prove an AI citation?

No. Growth in exact-brand queries is consistent with an AI citation but does not prove it as the cause. Compare weekly branded impressions against dated prompt-log entries and treat the result as directional.

What should you fix first when AI answers do not mention your brand?

Fix the first failing link in the chain. Check indexing and ranking first, then whether the page answers the question in a quotable passage, then whether the passage names the brand, and only then work on recommendation signals such as independent corroboration.

Is there a Google report for AI feature traffic?

Google's optimization guide points to the Generative AI performance report in Search Console for how people find content through generative AI features, and warns that third-party tools cannot access Google's internal metrics, so check their advice against official guidance.

Sources

  1. Google Search Central, Guide to optimizing for generative AI features on Google Search (opens in a new tab).
  2. Google Search Central, Creating helpful, reliable, people-first content (opens in a new tab).
  3. Measurement method on this site, How runs are recorded, and the downloadable Prompt Tracker (opens in a new tab).
Infographic: Ranking vs Retrieval vs Citation. Six outcomes that need six separate measures. Covers the short answer, What each outcome means, and how sure you can be, How to test AI answers properly, and Asfand's take.

A tall infographic of the key steps, to save on Pinterest or keep for later. Prefer to tap through? The same guide is a short Web Story.

Get the next note by email

Search Signals is the free newsletter version of this blog, with a two-page visibility checklist as a welcome gift.