How to Measure AI Citation Absorption

If your AEO dashboard reports one number, it is almost certainly citation count. And citation count is the wrong number.

A page can be picked by an engine, surfaced in the source list under an AI answer, and contribute nothing to the words the user actually reads. Another page can be picked half as often and shape every paragraph of the response. Citation count cannot tell those two cases apart.

A new measurement framework from researchers in late April 2026 finally names the gap. The paper From Citation Selection to Citation Absorption splits the AEO measurement question into two stages: was the page selected, and was its content absorbed. Across 602 prompts, 21,143 search-layer citations, and 18,151 fetched pages from ChatGPT, Google AI Overview/Gemini, and Perplexity, the authors show that selection and absorption move independently. Some engines cite broadly and absorb shallowly. Others cite less and absorb deeply.

If your reporting cannot distinguish those two motions, you are optimizing blind.

What citation absorption actually measures

Citation absorption is the share of selected pages whose specific language, evidence, or structure ends up in the AI's final answer.

Selection is a retrieval-layer event. The engine pulled your page into the candidate set, listed it in the sources box, and may have surfaced it to the user as a link. Absorption is an answer-layer event. The engine read your page, extracted a passage, and used that passage to construct what the user reads.

The two events can disagree in three ways. A page can be selected but not absorbed, a page can be selected and partially absorbed, and a page can be selected and fully absorbed with verbatim or paraphrased content reaching the answer. Citation count rolls all three into a single tick mark.

The frame in the funnel above is what one well-run client report should look like at the prompt level. Selection rate captures the retrieval funnel. Absorption rate captures what happens once your page is in the set. End-to-end share is the product of the two, and it is what most current dashboards report without separating the inputs.

Citation count tells you the wrong story

Once you have run the numbers separately, the gap between selection and absorption becomes the most useful diagnostic in AEO. Most B2B SaaS brands have a selection problem and an absorption problem at the same time, and the fixes are different.

A page that is cited but never absorbed is a page the AI used as decoration.

Reason #1: Citation count counts retrieval, not influence

The arxiv 2604.25707 study found that "citation breadth and citation depth diverge" across the three engines they tested. Perplexity and Google AI Overview list more sources per answer. ChatGPT cites fewer pages but the pages it does cite show "substantially higher average citation influence." Two engines can both report your page as cited and mean very different things by it.

Reason #2: Counts hide which pages are doing the work

A typical AEO dashboard shows your brand cited 23 times this week across a tracked prompt set. That number is the sum of three different motions: pages quoted verbatim, pages mentioned in the source list with no language reuse, and pages that influenced one supporting clause. Treating them as the same data point makes content prioritization arbitrary.

Reason #3: It rewards volume tactics that absorption ignores

Brands that publish high-volume thin content can move citation count without moving absorption. The May 2026 Google Core Update is already punishing this pattern in classic search. The day-six volatility data from the May 2026 rollout shows aggregator-style and thin AI-generated pages absorbing the heaviest negative impact. Absorption-aware reporting catches this drift early. Count-only reporting catches it after rankings fall.

Reason #4: It cannot distinguish a comparison page from a category footnote

The Aleyda Solis ecommerce AI citation study found that commercially valuable citations cluster on risk-reduction pages, reviews, comparisons, and returns or warranty pages. The same pattern shows up in our own work on B2B SaaS citation rate baselines, where comparison pages run well above category averages while feature pages sit below them. Those pages get absorbed deeply because they answer the buyer's risk question. A product feature page may be cited more often but get used only as a footnote. Count says the feature page is winning. Absorption says it is not.

Reason #5: Boards cannot defend a budget against a number with no behavior behind it

A CFO will ask one question about an AEO budget line: what did the citation actually do. Citation count tells you it happened. Absorption rate tells you how much of the answer your content shaped. The second number is the one budget conversations need.

The two-stage framework: selection vs absorption

The arxiv 2604.25707 framework defines the two events in plain operational terms.

Selection happens when the engine's retrieval layer hits a triggered search and pulls a candidate set, then narrows that set into the source list that surfaces under or alongside the AI answer. Selection is influenced by classic AEO signals: indexability, schema, canonical URLs, internal linking, brand authority, and the engine's own training-data priors about which domains to trust.

Absorption happens after selection. The engine reads the selected pages, extracts passages it can use, and constructs the answer. Absorption is influenced by passage structure, factual density, answer-block formatting, quotable one-liners, and the proximity of the relevant passage to the top of the page.

Selection asks:

• •Did the engine retrieve the page?
• •Did the page survive into the answer's source list?
• •Was the source list visible to the user?

Absorption asks:

• •Did the engine read content from the page?
• •Did that content reach the final answer's language?
• •Did the page shape the answer's structure or only decorate its citations?

Two different content interventions move these two events. Selection moves with structured data, content depth, and authority signals. Absorption moves with passage-level writing. A brand can fix selection and still lose absorption because nothing on the page extracts cleanly. We covered the writing side in detail in passages beat pages, which sits underneath any absorption strategy.

Step 1: Capture both selection events and absorption signals

The first job is to instrument the answer event so you can see both layers separately. Standard dashboards only show selection.

Run your tracked prompt set across ChatGPT, Perplexity, Google AI Overviews, and Gemini. Record the source list each engine surfaces and the URLs in it. That is your selection capture. Then store the full answer text alongside each event. The answer text is what you score absorption against. Tools that only return a citation count without preserving the answer body cannot give you absorption signal.

For brands without a vendor in place, this is a manual exercise. A weekly run of 30 to 50 priority prompts across four engines, with both the source list and the answer body saved, gives enough data to score absorption for the first time. The cost is two to three hours per week.

Step 2: Score per-citation absorption rate

For every selection event, ask one question: did the page's content reach the answer.

Three signals settle the question fast. The first is verbatim phrase match. If a sentence or distinct clause from the page appears in the answer body, the page was absorbed. The second is paraphrase with retained structure. If the answer reuses the page's framing, ordering, or data points without quoting verbatim, the page was absorbed. The third is unique fact attribution. If the answer states a number or claim that only your page makes in the visible source set, the page was the source.

If none of the three apply, the page was selected and not absorbed. Mark it as a selection-only event.

Aggregate across a prompt cluster to get the absorption rate. A page selected 17 times and absorbed 12 times has a 71% absorption rate. A page selected 30 times and absorbed 4 times has a 13% absorption rate. Both look identical in a count-only report.

Step 3: Segment absorption by page type

Once you have per-page absorption scores, group them by page type. This is where the diagnostic gets actionable.

Comparison and "X vs Y" pages typically post the highest absorption rates because the format invites quotable extraction. Documentation and how-to pages absorb well when the steps are short. Pricing pages often show high selection and low absorption because the engine does not quote price tables in answers. Feature pages and broad category pages tend to show low absorption because nothing extracts cleanly. Customer story pages and trust-and-compliance pages absorb well when the page leads with the headline outcome.

When a page type runs above 50% absorption, double down on volume. When a page type runs below 20%, rewrite the highest-traffic examples for passage-level extraction before publishing more.

Step 4: Compare absorption depth across engines

Engines do not absorb the same way. The arxiv 2604.25707 finding that ChatGPT shows higher per-page citation influence than Perplexity or Google means each engine demands a slightly different content posture.

A reasonable comparison table for a single client looks like this.

The same content investment will produce different reported gains on different engines. Reporting needs to separate them or the optimization signal averages out.

Step 5: Optimize content for absorption, not just selection

Selection optimization is what most teams already do. Schema, indexability, canonical hygiene, structured data, internal linking, page speed. These remain necessary. Absorption optimization is a different layer, sitting on top.

Three moves shift absorption the fastest.

The first is moving the direct answer above the fold. Engines absorb the passage that is closest to the heading that matches the query. A 40-to-60-word direct answer immediately under each H2 lifts absorption more than any schema change. We described the structural pattern in passages beat pages.

The second is writing one-liner pull quotes that are complete on their own. A short, declarative sentence that states a claim with no surrounding context will be quoted verbatim. A long compound sentence almost never is.

The third is publishing data and numbers in extractable form. "Citation absorption ran at 71% on comparison pages versus 13% on feature pages in our May 2026 sample" is absorption-friendly. "The data showed meaningful differences across page types" is not. The first version contains a number the engine can lift. The second version contains a claim with nothing to attribute.

What absorption-friendly content looks like

A short structural checklist captures the difference.

• •Each major heading is followed by a 40-to-60-word direct answer.
• •Specific numbers, percentages, and named studies sit in standalone sentences.
• •Page leads with the outcome or claim, not background or context.
• •Tables are short, with clear headers, and the data appears in surrounding prose as well.
• •One-liner pull quotes summarize each section's claim in 12 to 20 words.
• •Comparison content names competitors directly, with their positioning and the trade-off described in plain sentences.
• •Long pages get section anchors so passages near each heading can be cited independently.

Brands publishing in this format see absorption rates climb without selection rates moving. The page wins the same number of times it always did. It just does more work each time.

What absorption reporting does for a board deck

A board hearing AEO metrics for the first time needs three lines, not seven. Selection rate. Absorption rate. End-to-end share, with the previous period for comparison. We argued the broader case for separating reporting layers in the AI search measurement stack.

The story those three lines tell is the one a CFO can act on. "Our pages are selected by ChatGPT 17% of the time on tracked prompts, and 71% of those selections result in content reaching the answer. End to end, we shape 12% of the answers our buyers see. Last quarter that number was 8%. The 4-point lift came from a content rewrite on comparison pages." That is a budget defense. Citation count alone is a number with no behavior behind it.

For a more complete picture of how this slots into a wider measurement program, the seven CITE metrics framework covers Share of Model, Citation Rate, Recommendation Rate, and Citation Drift alongside absorption. Absorption sits underneath Citation Rate. It is the answer to "what happens after we are cited."

FAQ

What is the difference between citation count and citation absorption?

Citation count is the number of times an AI engine includes a page in the source list under or alongside an answer. Citation absorption is the share of those citations where the page's specific language, evidence, or structure actually appears in the answer body. The two move independently. A page can be cited often and absorbed rarely, or cited rarely and absorbed deeply. The arxiv 2604.25707 framework treats them as two separate measurement stages.

How do I calculate citation absorption rate?

For each citation event, check whether the page's content reaches the answer. Use three signals: verbatim phrase match, paraphrase with retained framing or data, and unique fact attribution. Mark the event as absorbed or selection-only. Divide absorbed events by total citation events to get the absorption rate. A page selected 17 times and absorbed 12 times has a 71% absorption rate.

Which AEO tools track citation absorption?

As of May 2026 no Tier-1 AEO monitoring vendor reports absorption as a defined metric. Profound, Peec AI, Otterly, Scrunch, Conductor, and Bluefish all report selection-side citation counts. Some preserve the full answer text, which makes manual absorption scoring possible. Until vendor coverage catches up to the arxiv 2604.25707 framework, absorption tracking is a manual or hybrid exercise layered on top of selection data.

Does citation absorption matter more than Share of Model?

They measure different things. Share of Model measures whether your brand appears at all in tracked AI responses. Absorption measures how much your content shaped the answers where you do appear. A brand with high Share of Model and low absorption is being mentioned without being read. A brand with low Share of Model and high absorption is the cited source whenever it shows up. Most B2B SaaS reporting today over-weights Share of Model and under-weights absorption.

How often should I score absorption?

Weekly at the prompt-cluster level for priority categories. Monthly at the page level across the full tracked set. Citation behavior shifts with model updates and core algorithm rollouts, and the May 2026 Google update plus Gemini 3.5 Flash going to default in AI Mode are both fresh reasons to keep the cadence tight. The pattern is the same one we documented in citation drift week to week: quarterly scoring misses too many shifts.

The 14% of B2B marketers currently measuring AI visibility, per Clearscope's 2026 SEO Playbook, almost all measure selection. Almost none measure absorption. The first team in any category to report both will have a measurement edge that compounds against competitors still reporting a single citation count.