Understanding Answer Engine Rankings

A search engine gives you a rank. An answer engine gives you a paragraph, and your brand is either the first thing named in it, one of several options, a footnote, or missing. Answer Engine Ranking is the measurement of which of those happened, per answer, per engine.

It is a different question from the visibility score, which folds position into a single index alongside mention rate and sentiment. This page is where you see the position on its own — and it is usually the faster route to a decision, because the distribution tells you what to change and the score does not.

Step 1: Open AI Rankings and confirm you are reading real data

AI Rankings sits in the Monitor group of the sidebar. It opens on your active project’s latest completed scan.

Before your first completed scan the page shows a clearly labelled sample: a preview table, a radar chart and a competitor bar chart with invented numbers. That is illustration of the layout, nothing more. Once a real scan lands, the page switches to your own data, the sample charts disappear, and four stat cards replace them — average position, queries tracked, number of first-place results, and mentions out of total results.

Read the average position card with its subtitle in mind: “when ranked”. It averages only the answers where you were given a position at all. It is not damaged by answers that never mentioned you, which means it can look excellent on a project with terrible coverage. It is a companion to the distribution below it, never a substitute.

Step 2: Learn the five buckets and where their boundaries fall

Every analysed answer is sorted into exactly one of five buckets, and the boundaries are fixed:

The fourth row is the one to internalise, because two quite different things land in it. An answer that lists twelve tools and puts you eleventh is fine-print. So is an answer that mentions you in passing — “unlike the approach taken by your brand…” — where there is no list and therefore no position to hold. Both are weak outcomes, but the second is often a content signal rather than a ranking one: the engine knows you well enough to reference you and did not consider you a candidate answer to the question asked.

Step 3: Read the distribution before the average

The Answer Engine Ranking card shows all five buckets as counts, then lists every answer underneath with its engine, prompt and bucket. Read the counts first.

The shape tells you more than any single number:

• Weighted toward first and prominent — you are a default answer in this category. Protect it; watch for drift when engines update.
• A bulge in mid-list — you are a known option that never wins the recommendation. This is the most common shape, and the most actionable.
• Heavy fine-print — check whether those are late list positions or unranked passing mentions; they need different responses.
• Heavy absent — this is a coverage problem, not a ranking one. No amount of positioning work moves an answer you do not appear in at all.

Position matters for the same reason it matters in a human recommendation: the first name someone hears is the one they remember, and a brand described in its own sentence lands harder than one dropped into a list of six. Ranking each answer rather than just recording a mention is what lets you see that difference.

We deliberately publish no percentage here. Numbers of the form “first mention drives N% higher recall” circulate widely in AEO writing without a traceable source, and inventing one would undercut the point of measuring at all. What your own scan history gives you is the honest version: the position distribution for your prompts, moving over time as you change what you publish.

For what it is worth, the position effects that are well documented in the literature are on the input side, not the output side — Liu et al. (2023) showed that models attend unevenly to material depending on where it sits in their context window. That is a finding about how models read, not about how people react to being named third in an answer, and extrapolating one to the other would be exactly the kind of borrowed statistic this page refuses to print.

Step 4: Work the query table engine by engine

Below the distribution, Query Rankings groups results by prompt. Each row shows the question and then one chip per engine, carrying whichever of two things that answer produced: a #n position badge when the engine named your brand somewhere orderable, or the band label — Fine-print for a mention with nothing to order it against, Absent for an answer that did not name you at all. Hover either label for the sentence explaining it.

Sentiment sits alongside, and only on the chips that were mentioned. That is deliberate rather than an omission: sentiment is measured toward your brand in that answer, so an answer that never named you expresses none, and the stored neutral on those rows is a column default rather than a reading. An absent chip with a sentiment beside it would be inventing a measurement.

This is where disagreement between engines becomes visible, and disagreement is the useful signal. The same question answered by seven surfaces routinely produces four different shortlists, because each engine draws on different sources and, in the case of grounded engines, a live web search performed at that moment. Google’s own guidance on AI features describes these surfaces as answering conversational questions rather than returning a ranked list, which is precisely why a per-engine chip row is the right shape for this table.

A prompt where one engine puts you first and three do not mention you at all is a much sharper brief than an averaged score: something one engine has access to, the others do not.

Step 5: Tell a genuine absence from a missing measurement

Not everything that looks like a zero is one.

• Answers whose engine call failed are excluded entirely from this page. They are not counted as absent, because a timed-out request is missing data rather than evidence you were left out.
• An engine whose every call failed is dropped from the scan’s per-engine reporting, for the same reason — an all-zero row is indistinguishable from genuine invisibility.
• Google AI Overviews only appears when a SERP provider is configured for the deployment. If your rows never include it, that is a configuration state, not a ranking of zero.

So before reacting to a wall of “absent”, confirm the answers actually came back. The Activity page lists every scan job and its outcome, which is where a partially failed run shows up.

Step 6: Decide what to change

Map the bucket to the work:

• Mid-list → prominent. The engine already treats you as a candidate; it is not treating you as the best fit for the question asked. Publish content that answers that specific question directly and states who the answer is for, rather than a general product page.
• Fine-print with no rank → any rank. You are referenced, not recommended. Comparison and alternatives pages tend to move this one, because they put you into the answer as an option instead of as context.
• Absent → present. Treat it as a content gap and work it as one.
• Prominent → first. The slowest and least controllable move. It usually follows category-level authority and third-party corroboration rather than anything on your own site — see Citation Intelligence Deep Dive for which outside sources the engines are actually leaning on.

Re-measure on the same prompt set. Changing prompts and position at the same time means you cannot attribute the move to either.

Step 7: Where to go next

• Answer engine ranking — the feature page behind this view.
• What is answer engine ranking? — the concept, without the product specifics.
• Understanding position drift — why a position moves when you changed nothing.
• Understanding Your Visibility Score — how position folds into the headline number, and why the two can disagree.