AEO Metrics That Matter: Mentions, Citations, Split

Most AEO reports have one number and a green arrow. The number moved 0.5 points. Someone calls it a win. Next month it moves back.

AEO metrics only help if you know what they count, which engine they came from, and whether the move survived noise. Mentions, citations, and engine split are the three you can defend. Everything else is optional until those three are stable.

I pulled two sources that take this seriously: Discovered Labs on KPIs and test-bed noise, and Brainlabs on what sits beyond mentions. Then I mapped that to what we actually store in AI SEO Tracker.

In this guide, you'll learn:

• Mention rate vs citation URL
• Engine split, the metric people skip
• Share of voice without the vanity
• What not to treat as a KPI yet
• A reporting stack that survives a CFO

We'll break down each measure below.

Mention rate vs citation URL

Start with two counts. Do not blend them.

Mention rate = answers that name your brand / answers you sampled.

The brand string (or an alias) appears in the answer. There may be no link. Peec's docs call this Visibility. Same math. Discovered Labs often says "citation rate" when they mean "how often you appear." I will not do that. Appearance without a URL is a mention. A URL is a citation.

Citation URL rate = answers that attribute a URL on your domain / answers you sampled.

Perplexity puts numbered sources under the answer. ChatGPT sometimes links, sometimes does not. AI Overviews show citation cards. If you only count named brands, you miss pages that got used as evidence. If you only count URLs, you miss the shortlist that never links out.

Peec splits this as brand visibility vs source visibility. You can be a source without being named. You can be named without being a source. Both gaps are real:

• Named, never cited: the model knows the entity. It does not trust a page of yours enough to show a URL.
• Cited, never named: a guide on your domain got retrieved. The brand is weak.

HubSpot's AEO note is the same split: mentions build association, citations can send a session you can see in analytics.

Write both on the slide. Never one.

How to compute mention rate without cheating

Denominator = prompts in the frozen set × engines × runs you keep.

Numerator = those answers where an alias matches.

Rules:

1. Use unbranded prompts for the headline rate. Branded prompts inflate the number. Discovered Labs saw a 10.8% reported rate collapse to 1.9% once brand-anchored prompts were removed.
2. Count an answer once. Pronouns after the first name are not extra mentions.
3. Store the answer. If you cannot reopen it, you cannot defend the count.
4. Keep a written alias list. Misspellings count if a human would still see you.

A 20-prompt, one-run, one-engine "rate" is a sketch. Treat it as a sketch.

Engine split, the metric people skip

ChatGPT, Perplexity, Gemini, Copilot, and AI Overviews do not share a retrieval stack. Averaging them is how you hide a dead engine.

Brainlabs' Consensus Position is the useful version of this. They treated each platform as a judge. For each prompt they asked: zero engines, some engines, or all engines. In their 13-week test (33 prompts, four platforms), 18 prompts showed the brand on none of the four. Six showed it on all four. Nine sat in the middle. The average (about 1.3 platforms per prompt) looked dull. The distribution told them where to work.

That is the report I want:

Prompt	ChatGPT	Perplexity	Gemini	Copilot	AI Overviews
Best X for Y	Named	Cited URL	Miss	Miss	Miss
X vs Z	Miss	Named	Named	Miss	Cited

Three strategies fall out:

• Zero engines: you have no consensus. You need third-party mentions and a page that answers the prompt in the first two sentences. See LLM SEO.
• One or two engines: close the gap. The brand is knowable. A specific source or format is missing on the engines that skipped you.
• All engines: defend it. Do not "optimize" a winning prompt into a different article.

Profound's Index also breaks industries by LLM. Peec filters by model. AI SEO Tracker shows the engine split on purpose. A single "AI visibility %" is a compromise for people who will not read a table. Do not let it be the only line in the deck.

Perplexity in particular is a citation engine. ChatGPT is a mention engine. If you average them, you will think you "have citations" because Perplexity linked you, while ChatGPT still never says your name. Read how Perplexity picks sources before you set a citation target that ChatGPT cannot hit.

Share of voice without the vanity

AI share of voice = your mentions (or citations) / all tracked-brand mentions (or citations) in the same answer set.

Peec's definition is the clean one: percentage of brand mentions compared with the brands you chose to track. High SoV means you are the main name in those chats, not that you own the internet.

Two versions, keep them separate:

Flavor	Formula	What it is good for
Mention SoV	Your names / all tracked names	Shortlist share
Citation SoV	Your URLs / all cited URLs	Source authority

A 25% mention SoV and a 8% citation SoV means people (and models) say your name, but they prove the claim with someone else's URL. That is a content problem. The reverse means you are a footnote, not a brand.

Do not compare your SoV in tool A to a competitor's screenshot from tool B. Prompt sets, engines, and match rules differ. Discovered Labs and every honest roundup say the same thing: pick one tracker and read it as a trend line, not as physics.

For the older media version of this metric, see share of voice measurement. The AI version uses answers as the inventory, not ad slots.

What Brainlabs adds once the basics work

After mentions and citations exist, Brainlabs tested five deeper KPIs on live data. They say these are not industry standards yet. Treat them as extra questions, not as a new dashboard theme.

1. Consensus Position. Covered above. Distribution across engines, not an average.

2. Share of Recommendations. Named in a list of eight is not "the answer." In their test: 42 mentions, 4 that met a recommendation bar (lead choice or explicit suggestion). 12.1% share of recommendations. Presence and advocacy are different jobs.

3. Citation half-life. How long a cited URL stays in the set. Their median among URLs cited more than once was 28 days. 76% of cited URLs appeared one week and never again. A one-week citation is a spike. A URL that lasts is an asset.

4. Share of Narrative. Does the model describe you with the attributes you chose? Visibility can be high and still wrong ("cheap" when you sell "precise"). Score 5 to 8 positioning attributes every week. Track the gap, not a fake universal score.

5. Prompt space coverage. Are you measuring demand you care about? Their 33 prompts missed huge GSC themes. A 38% mention rate on a set that covers 61% of demand is not a 38% mention rate on the market.

They also flag source concentration: if two domains drive most of your citations, one robots.txt change can wipe the row. Useful as a risk note, not as a score to "optimize down."

Use these after mention rate, citation URL, and engine split are in the weekly report. If you cannot compute mention rate cleanly, you cannot compute half-life.

What Discovered Labs insists you bound

Their test-bed post is the other half of adult measurement.

Same prompt, same model, different output. Temperature and GPU batching do that. Your prompt list is also a sample of an unknown buyer-query population. So every rate has two noise sources.

Their buyer checks, shortened:

On the prompt set

• Sampled from real buyer language, not 20 vanity questions
• Surface form (length, formality) not driving the variance
• Anchored vs unanchored split published

On the variance

• An interval on every snapshot, not a lonely 12.4%
• Sample size large enough that a 0.5 point move is not the whole interval
• Between-prompt vs within-prompt noise separated if you can resample

On the trend

• Evidence that "it moved" beats "it stayed flat"
• The new rate holds across several snapshots, not one lucky Monday
• Weekly as the default. Daily only after a launch.

You do not need a Bayesian sermon in a client Slack. You do need to stop calling a 12.4% to 12.9% bump a win on 40 prompts.

We run weekly scans in AI SEO Tracker for that reason. Daily feels productive. Weekly is closer to a shift you can act on.

A reporting stack that survives a CFO

One slide. Four lines. Engine table in the appendix.

1. Mention rate (unbranded), this week vs last, per engine
2. Citation URL rate, same window
3. Share of voice vs the 3 competitors you actually lose deals to
4. Top cited domains in the answers you lost (so PR and content have a list)

Optional fifth line once you have 8+ weeks: prompts that flipped from miss to mention and stayed there. That is half-life in plain language.

Do not lead with:

• A 0 to 100 "AI score" with no formula
• Organic sessions (different channel; pair it, do not replace the AEO lines)
• Rankings. Ahrefs has shown that most ChatGPT citations are not Google top 10 URLs. Rankings still help AI Overviews more than ChatGPT. See AI SEO vs traditional SEO.

Pipeline from AI referrals is real when you can tag it (ChatGPT UTMs, Perplexity referrer). Discovered Labs wants that as KPI three. Fine, as a companion. It is not a substitute for mention rate. Zero-click shortlists will never show up as sessions.

For traffic fingerprints in Search Console, use track AI search in GSC. That is retrieval evidence. It is not mention rate.

Worked example (small set, honest math)

25 unbranded prompts. Five engines. One weekly run. 125 answers.

• ChatGPT named you in 4 of 25
• Perplexity named you in 6 of 25 and cited your URL in 5 of 25
• Gemini named you in 2 of 25
• Copilot 1 of 25
• AI Overviews 3 of 25, two of those with a citation card

Mention rate overall: 16 / 125 = 12.8%. Say that, then immediately show the split. ChatGPT 16%. Perplexity 24%. Gemini 8%. Copilot 4%. AI Overviews 12%.

Citation URL rate overall: 7 / 125 = 5.6%. Almost all Perplexity and Overviews. ChatGPT contributed mentions without URLs.

If a competitor was named in 40 of 125, your mention SoV among is 16 / (16+40) = 28.6% of the two-brand set, not 12.8% of the market. Always say who is in the denominator.

That is a report. A sparkline labeled "Visibility +4%" is not.

What to do when a metric moves

Move	Likely cause	First check
Mention rate up, citations flat	Name recognition, still no quotable page	Page structure, facts block, Page Inspector
Citations up, mentions flat	A guide got retrieved, brand still weak	Entity consistency, third-party listicles
One engine jumps, others flat	That engine's index or a source it likes	The citation URLs on that engine only
All engines drop together	Prompt set change, alias bug, or a category rewrite	Did we edit prompts? Did a legal name change?
±2 points on 25 prompts	Noise	Wait two more weeks

If you need a monitor that already stores mention rate, citation URL, and engine split, that is the job AI SEO Tracker was built for. $49 for the first audit. From $79/mo if the weekly trend is worth paying for.

AEO is not a new religion. It is a measurement problem with sloppy denominators. Fix the denominator. Split the engines. Keep mentions and URLs on different rows. Then you can talk about half-life and narrative. Not before.