---
title: "AEO Benchmarks: Which Citation-Rate Numbers You Can Quote"
description: "Most published AEO benchmarks come from vendors and state no sample. Here is a short test for any figure, applied to five popular ones, plus what to benchmark against instead."
author: "Team"
category: "Search & AI Visibility"
date: 2026-11-12T09:25:02.061Z
canonical: "https://contentagents.dev/blog/aeo-benchmarks-claim-check-citation-rates-volatility-and-tiers"
---

# AEO Benchmarks: Which Citation-Rate Numbers You Can Quote

![checklist, check, confirm](https://cdn.pixabay.com/photo/2019/06/17/08/25/checklist-4279388_1280.jpg?w=1200&q=75)

> Most published AEO benchmarks come from vendors and state no sample. Here is a short test for any figure, applied to five popular ones, plus what to benchmark against instead.

Open any AEO guide this year and you will find a tier ladder: under so many percent is "early", over some other number is "leading". The ladder is tidy. The data behind it is usually missing. This piece gives you a four-question test for any benchmark, runs five popular figures through it, and ends with a recommendation that will annoy anyone who wants a number for a slide.

It is written for the person who has to set a target or defend a report on AI visibility. Several of the figures below come from AirOps, which sells an AEO platform, so its data carries a commercial interest. Content Agents sells content software and has an interest of its own. Nothing here was re-measured; the verdicts are about what each page lets a reader verify.

## Four questions that sort a benchmark from a slogan

- **Who counted, and what do they sell?** A tracking vendor benefits when the market feels unmeasured.

- **What was the sample?** Prompts, engines, brands, and the date range. Missing any one makes the figure unusable as a target.

- **Is it a rate or a rule of thumb?** A measured average is data. A threshold labelled "leading" is an opinion with a decimal point.

- **Does an independent source point the same way?** One dataset repeated by ten blogs is still one dataset.

## The five figures, in the order readers meet them

### Volatility: "only 30% of brands stay visible"

AirOps' audit checklist says about 30% of brands remain visible from one AI answer to the next, and about 20% across five consecutive runs, attributing it to its own research ([AirOps audit checklist](https://www.airops.com/blog/aeo-audit-checklist)). Its metrics article repeats the 30% and adds that about 57% of brands that vanished came back within two runs ([AirOps metrics](https://www.airops.com/blog/aeo-metrics)). Neither page states how many prompts, engines or brands sit behind the number.

The direction is believable, because Google says its AI features may run several related searches behind one answer, which adds variation by design ([Google Search Central](https://developers.google.com/search/docs/appearance/ai-features)). The exact value is not checkable. Treat it as a reason to repeat each prompt, not as a baseline.

### Position: "60% of citations come from outside the top 20"

The checklist credits AirOps' State of AI Search report for the 60% figure. The report itself was not read here, only the article that repeats it, and another AirOps article cites a different split, 38% of AI Overview citations from top-10 organic results, with a link that lands on a general page ([AirOps brand mentions](https://www.airops.com/blog/ai-brand-mentions)). The two figures answer different questions, and neither can be traced to a method.

Do not tell stakeholders that rankings stop mattering, or that they still decide everything. Measure the pages you own.

### Structure: "headings give 2.8x citation likelihood"

The checklist says pages with clean, sequential headings are 2.8 times likelier to be cited ([AirOps](https://www.airops.com/blog/aeo-audit-checklist)). No control group is described. Well-edited pages tend to have clean headings and many other virtues, so the multiplier cannot be pinned on headings. The advice is still cheap and sound. Fix the headings; skip the multiplier.

### Tiers: "a citation rate above 15% is leading"

The same vendor lists citation-rate bands from early to leading, and says they vary by category. No source is given for the cut-offs. A second tier ladder, circulated by an ecommerce agency in June 2026 and credited to a vendor blog, puts a "leading" category-query rate at 25% and above, with no sample size either ([AI Advantage Agency](https://aiadvantageagency.com/ecommerce-aeo-benchmarks/)). Two ladders, two thresholds for "leading", zero shared method. That disagreement is the finding.

### Unbranded mentions: "0.34% on average"

AirOps reports a 0.34% average mention rate and a 5.42% citation rate for one target unbranded prompt ([AirOps](https://www.airops.com/blog/ai-brand-mentions)). This one holds up as an illustration: unbranded prompts rarely name any brand, and citations outnumber mentions. It is one dataset on one prompt type, so read it as an example, not an industry average.

## What a stated sample actually looks like

One vendor study did publish its numbers. Digital Applied reviewed a tracking dataset of 2,437,107 citation records across 28,725 domains in a single vertical, Australian insurance, covering seven non-consecutive months from late 2025 to mid 2026 ([Digital Applied review](https://www.digitalapplied.com/blog/ai-search-citation-decay-study-what-it-measured)). Its headline: 57.2% of domains were cited in one sampled month and not seen again, and only 2.7% appeared in all seven.

The same review ([Digital Applied](https://www.digitalapplied.com/blog/ai-search-citation-decay-study-what-it-measured)) also lists the holes. Monthly volume in the data swung roughly 29x, which the vendor puts down to a changing query set. The survival table sums to fewer domains than the headline total. No raw data or query list is public. This is what a better benchmark looks like and it still falls short, which tells you how thin the field is. Its value for you is the shape of the finding: AI citations churn, so a single monthly snapshot misleads.

## Skip the tiers, keep a baseline of your own

Here is the position. A tier ladder is the wrong instrument for a team under two years into AEO, and the better move is to retire it from your reporting entirely.

The reasoning comes straight from the table above. Two public ladders disagree on what "leading" means. Neither states a sample. Your rate also depends almost entirely on how you wrote the prompt list, which is why a fair set matters more than the target (see [how to build a prompt set that is not rigged](/branded-vs-unbranded-prompts-build-a-prompt-set-that-is-not-rigged)). A tier built on someone else's prompts will tell you that you are "developing" or "strong" for reasons unrelated to your market.

What replaces it is cheaper and harder to argue with. Fix the prompt list. Run each prompt several times, for the reasons in [why AI answers change every run](/why-ai-answers-change-every-run-and-how-to-measure-around-it). Record your rate and the rate of three named competitors on that same list. Report the change against your own last quarter, then use [share of voice](/share-of-voice-in-ai-search-definition-formula-and-a-worked-example) to compare against rivals.

Reproduction would settle it: if a vendor publishes prompts, engines, brands and dates for a tier ladder, and an independent team reproduces it. Until then, a vendor tier belongs in a footnote labelled "vendor-reported", never in a target.

## If a leader insists on a ladder

Some executives will not accept "we compare against ourselves". Give them a ladder, but build it from your own data: take your last four quarterly rates on the fixed prompt list, call the lowest "baseline" and the highest "best so far", and let the middle two define the range. It is arbitrary in the same way the vendor ladders are, with one difference. Every cut-off traces to a measurement somebody can inspect, and it moves only when the prompt list is deliberately changed, which you then log as a break in the series.

Keep the competitor view next to it. A rate that rises while three rivals rise faster is a loss, and no tier label will tell you that.

## One new source worth adding: Search Console's own AI impressions

If your benchmark question is "how often does Google show my pages in AI features", you no longer have to infer it from prompt sampling. Google announced Search Generative AI performance reports on 3 June 2026, covering AI Overviews, AI Mode and generative features in Discover, and said they reached all websites worldwide by 31 August 2026 ([Google Search Central](https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports)). They show impressions by page, country, device and date, with no clicks or position yet.

That is first-party, per-page, and free, so it works as the anchor against which a prompt-sampled citation rate can be sanity-checked. Note that Google's older AI features page, last updated in December 2025, does not mention these reports, so do not use it to decide whether they exist. For a fuller measurement set, read [the complete AI visibility measurement guide](/ai-visibility-metrics-complete-measurement-guide).

## Before a vendor figure reaches a board deck

- Label it "vendor-reported" and name the vendor.

- Write the year beside it. Several of these figures are from early 2026 and the engines change monthly.

- State what is missing: sample, engines, dates.

- Pair it with your own number from the same quarter.

## FAQ

### Are there reliable AEO benchmarks?

Few. Most come from vendors and omit the sample. The most transparent dataset found covers one vertical, so use your own history as the baseline.

### What is a good AI citation rate?

It depends on your prompt set and category. Published tier ladders disagree with each other and give no method, so compare yourself with your own past results and named competitors.

### Can Search Console show AI Overview visibility now?

Yes. Since June 2026 Google has offered Search Generative AI performance reports with impressions by page, country, device and date. They show no clicks or position.

### Should I put vendor statistics in a board report?

Only labelled as vendor-reported, dated, and with the missing method noted, alongside your own measured numbers.


---
Source: https://contentagents.dev/blog/aeo-benchmarks-claim-check-citation-rates-volatility-and-tiers