What should I measure for AEO beyond citation counts?

Track the mention rate on buyer-intent prompts, share of voice against a fixed competitor set, the accuracy of what engines say about your brand, your position in the answer, AI referral sessions and their conversion, branded search lift, and self-reported attribution. Collect each on a locked prompt set with repeated runs, and read the citation count as one input among these rather than the headline figure.

Is citation count a vanity metric in AEO?

Not entirely. A citation count is an early, partial signal: it shows that an engine used one of your pages as a source. It becomes misleading when it is the headline KPI, because it treats informational and buyer-intent prompts the same, ignores mentions without links, and is often read from a single run of answers that vary from run to run.

Why does my AI visibility score change even when my citations stay flat?

A composite visibility score combines several inputs, and most of them move independently of the citation tally. A change in your mention rate, in the intent mix of the prompt set, in your position inside the answer or in a competitor's share will shift the score. Run-to-run variance moves it too. Publish the weights with the score so a method change is never read as a visibility change.

What is a good citation rate or share of voice in AI search?

We found no independent, non-vendor benchmark of a good citation rate or share of voice by industry. The useful reference point is your own competitor set, measured on the same prompts, engines, locale and schedule, and your own trend over repeated runs. A rising share against the same competitors means more than any absolute figure taken out of context.

How many prompt runs do I need before an AEO metric is reliable?

There is no fixed number. A mention rate is a proportion, and its margin of error shrinks as the number of prompt-run pairs grows, so a small sample gives a wide range and a large one a narrow range. Lock the prompts, engines, locale and schedule before comparing periods, and give noisy signals such as position and sentiment more runs than simple presence.

AEO Signals to Measure Beyond Raw Citation Counts

Beyond raw citation counts, measure mentions on buyer-intent prompts, share of voice against a fixed competitor set, the accuracy of what engines say about you, your position in the answer, AI referral sessions and branded search lift. Each signal needs repeated runs before it can be trusted, and none of them proves revenue on its own.

Why a raw citation count is a weak AEO signal on its own

An answer engine is a system that composes an answer instead of returning a list of links. Rankings, clicks and organic sessions describe performance in a list of results; they do not show whether your brand appeared inside the answer that replaced it. Pew Research Center reported on 22 July 2025 that Google users who encountered an AI summary clicked a traditional result link in 8% of visits, against 15% for those who did not.

Visibility inside the answer is broader than citation. A citation is a linked source; an engine can also name your brand, recommend it or describe it without linking to you (see how an AI mention differs from an AI citation). Showing up in ChatGPT can therefore mean being named, recommended, cited or described correctly. A report that collapses these into one tally answers none of them.

A raw tally also weighs every prompt the same. A citation on an informational prompt such as "what is schema markup" counts as much as a mention on "best CRM for a small agency", where the reader is choosing a vendor. The tally can grow while the brand stays absent from every prompt that leads to a purchase.

And a single run is a sample, not a measurement, because AI answers are non-deterministic. SparkToro research, dated 28 January 2026, estimates less than a 1 in 100 chance that ChatGPT or Google's AI, asked 100 times, gives the same list of brands in any pair of responses.

The AEO signals worth measuring beyond citations, compared

Each row is one signal. Share of voice, often confused with citation count, is the share of answers naming your brand on a locked prompt set, compared with a fixed competitor set measured on the same prompts and runs. It moves when a competitor gains ground, even if your counts stay flat. Noise ratings are our inference, not published measurements.

Signals measured inside the answer

Signal	What it measures	How to collect it	Run-to-run noise	What it can prove
Mention rate on buyer-intent prompts	Share of runs naming the brand on buyer prompts	Repeated runs of tagged buyer prompts	Medium	Presence where vendors are chosen, not purchases
Share of voice	Your mentions relative to a fixed competitor set	Same locked prompts and runs for every brand	Medium	Relative standing against named competitors
Citation rate	Share of runs linking one of your URLs as a source	Parse the sources each engine displays	Medium	Use as a source, not persuasion
Answer position or prominence	Where the brand sits: first, in a list, in passing	Record order and framing per run	High	Prominence, only as an average over many runs
Accuracy of brand facts	Share of stated facts that are correct	Score answers against a verified fact sheet	Medium	Whether engines describe you correctly
Sentiment	Tone of the description	Classify wording per answer, with human checks	High	Direction of perception, read over many runs
Prompt coverage	Share of prompts where the brand appears in at least one run	Aggregate presence per prompt	Low	Breadth of topics where the brand exists

Signals measured outside the answer

Signal	What it measures	How to collect it	Run-to-run noise	What it can prove
AI referral sessions and conversion	Visits from AI referrers and what those visitors do	Custom channel group plus conversion events	Low, not run-based	That answers send visitors who act, undercounted
Branded search lift	Change in searches for the brand name	Search Console queries containing the brand, over time	Medium	Correlation with visibility, not cause
Self-reported attribution	What buyers say led them to you	Form field and sales questions	Medium	Influence analytics cannot see, on small samples

AI referral sessions and the conversion quality of those sessions carry more business value than citation volume, because they record a visitor who acted. They also undercount, as the pipeline section explains.

Content structure and specificity sit behind these signals, not among them. We infer that specific pages with checkable facts are easier for an engine to use and to describe accurately. Structure is the lever; the signals tell you whether pulling it worked.

Signals also differ by engine. Google Search Central states, in documentation consulted on 12 November 2026, that "sites appearing in AI features (such as AI Overviews and AI Mode) are included in the overall search traffic in Search Console", so those clicks are not reported on a separate line. We could not verify how OpenAI and Perplexity document source display, so check it per engine before comparing citation rates.

We found no independent, non-vendor benchmark of a good citation rate or share of voice by industry. The reference point is your own competitor set, on the same prompts, over time.

How to weight a buyer-intent mention against an informational citation

AI-referred visits and buyer-intent visibility deserve more weight than raw citation volume, because they sit nearer to a decision. The weights are a choice, not a law: defend them, then keep them stable.

Hypothetical worked example. Tag each prompt as informational, consideration or decision, and give decision prompts the heaviest weight and informational prompts the lightest. Score mention and citation separately, because being named and being linked are different outcomes. Brand A is cited on many informational prompts and named on almost no decision prompts: its raw count looks strong, but its weighted composite is low, because the heaviest prompts contribute almost nothing. Brand B has fewer citations and steady mentions on decision prompts, and scores higher. The composite matches what a buyer sees; the raw count does not.

The same mechanics explain why a visibility score can move while citations stay flat. A shift in the intent mix of the prompt set, a move from a passing mention to first in the answer, or a change in competitor share all change the composite and leave the citation tally untouched.

Publish the weights next to the score in every report, and recompute earlier periods when a weight changes, so a change in method is never mistaken for a change in visibility.

How many prompts and runs before a signal is reliable

The same prompt, run again, can return a different answer. OpenAI's API documentation, consulted on 12 November 2026, states that "Chat Completions are non-deterministic by default (which means model outputs may differ from request to request)." Trends across repeated runs therefore matter more than any single snapshot.

A mention rate is a proportion: each prompt-run pair is one trial, and the brand appears or does not. Its margin of error shrinks with the square root of the number of pairs, so quadrupling the pairs roughly halves the range.

Hypothetical example: a brand appears in about half its runs. On a handful of prompt-run pairs, the plausible true rate spans most of the scale and a weekly jump means nothing; on a sample many times larger, the range narrows to a few points either side.

Before comparing periods, fix the prompt set, the engines, the locale and the run schedule. Log dated model and product changes, so a broken trend is traced to its cause, not credited to your content.

Noise varies by signal. Whether the brand appears at all is a binary outcome and settles quickly. Answer position and sentiment wording vary more, because they depend on how the engine orders and phrases the whole answer. Give them more runs per prompt, and report them as averages with their range.

How to audit the accuracy of what AI engines say about your brand

Presence, accuracy and sentiment are separate signals. An engine can name you on every buyer prompt and still quote an old price or a feature you removed. A repeatable audit:

Write a dated reference sheet of verified brand facts: pricing model, core features, markets, founding facts.
Fix a set of brand prompts and comparison prompts against named competitors.
Score each answer per fact as correct, wrong, outdated or missing.
Compute the error rate over repeated runs: wrong plus outdated facts, divided by facts checked.
Score sentiment separately, as tone rather than fact.

Set an internal escalation threshold, for example any wrong price or wrong capability on a decision-stage prompt. The threshold is your team's choice: we found no published standard for it.

A wrong fact is not a wrong citation, which is a source attached to a claim it does not support. In a study published on 6 March 2025, "AI Search Has a Citation Problem", the Tow Center for Digital Journalism at Columbia Journalism Review reports that the AI search tools it compared collectively "provided incorrect answers to more than 60 percent of queries." Read it as a reason to audit, not as an expected error rate for your brand.

Which signals connect to pipeline, and what correlation cannot prove

AI-referred sessions and their conversion rate sit closest to business value, but many AI-influenced visits arrive as direct or branded traffic: a buyer reads an answer, then types your name later. For the referrals you can see, create a custom channel group. Google Analytics Help, consulted on 12 November 2026, shows channels defined by source conditions such as "Source exactly matches"; apply them to referrers such as chatgpt.com and perplexity.ai. Its full setup is covered in setting up an AI referral channel in Google Analytics.

Ahrefs reported on 16 June 2025 that "AI search traffic accounts for just 0.5% of our traffic for the last 30 days", and its article title states that this traffic drove a far larger share of signups. That is one site and one window: a single-site observation, not a gain to expect.

Branded search lift and direct-traffic lift are proxies. If buyer-prompt visibility and branded queries rise together, they are correlated; a campaign, press coverage or seasonality can move both, so correlation does not show cause.

Self-reported attribution is a complement that analytics cannot replace. Add a "how did you hear about us" field with an AI assistant option, and ask sales which prompt a prospect used.

We found no published evidence demonstrating a causal link from an AEO signal to pipeline. A test would need a holdout set of prompts or pages, a dated baseline and a change in one variable. For a method that reports the effect as a range, see connecting AI visibility to revenue as a range.

Where to start this week: build a signal scorecard

Tag your existing prompt set by intent tier.
Choose a handful of signals from the tables, inside and outside the answer.
Fix engines, locale and runs per prompt.
Write the brand fact sheet and score one accuracy audit.
Set up the AI channel group and a self-reported attribution field.
Record a dated baseline for each signal.

Lay the scorecard out with one row per signal and these columns: signal, definition, weight, baseline value, run count, date.

To set your first baseline, run a free Namedrop scan and see which engines name your brand on your buyer prompts and which competitors they name instead.