AI Search answer-state taxonomy: classify the answer before fixing citations

Method

Why mention rate and citation rate are not enough

AI visibility dashboards often start with two useful fields: did the answer mention the brand, and did it cite the domain? That is the right beginning, but it is not enough for operations.

Frase defines AI search tracking as monitoring how a brand, content, and expertise appear across AI-powered search engines, including citation frequency, brand mentions, competitive performance, topic gaps, and citation quality ². Otterly.ai similarly separates linked citations from text-only mentions and recommends tracking prompt libraries, competitive benchmarks, source URLs, and historical trends ³. Profound's product language adds source categories such as owned, competitor, earned media, PR wire, social, and institution ⁴.

Those are useful measurement dimensions. The missing layer is the repair decision. A brand can be mentioned without being cited. A page can be cited without shaping the answer. A competitor can be cited for a claim that your first-party page was never built to support. A third-party review can be the correct source type for a trust prompt. A hallucinated claim can look like visibility until someone checks the source.

That is why the unit should be an answer state, not only a visibility score.

The Tow Center's research on generative search makes the risk visible. In a 2025 test of eight AI search tools, the team found that systems often answered inaccurately, failed to cite original sources, cited syndicated copies, fabricated or broken URLs, and sometimes answered with confidence when they should have declined ¹. For a marketing team, the implication is direct: "cited" is not a single positive state. Some citations are correct, some are wrong-source, some are stale, and some are not actually used by the answer.

Method

The answer-state taxonomy

Use this table as the first pass before making any content decision.

The table is intentionally blunt. A weekly operator should be able to label a row in less than a minute. A strategy review can add nuance later.

Method

How to map states to owners

An answer state is useful only when it creates the right owner.

If the state is absent, the owner might be technical SEO or content strategy. First prove the page exists, is crawlable, is canonical, appears in sitemap/feed/LLM discovery surfaces, and maps to the prompt. If none of that is true, the problem is not "AI hates the page." The page may not be eligible or relevant.

If the state is mentioned but not cited, the owner is usually content plus source architecture. The brand is in the model's answer space, but the source graph is weak. The repair is an extractable source unit: direct answer, definition, comparison, data point, FAQ, schema parity, and internal links from related pages.

If the state is cited but not absorbed, the owner is editorial. The page is visible enough to be cited, but the answer does not reuse the page's distinctive language or proof. Add stronger answer units near the top: a short definition, a table, a numbered process, a claim with source markers, or a comparison row. The goal is not longer prose; it is easier extraction.

If the state is competitor cited, the owner is competitive intelligence plus ContentOS. Compare the competitor's source against the desired page: evidence type, recency, source category, page structure, independent corroboration, and claim specificity. The next task may be a page section, a new page, a source pack, or a third-party corroboration push.

If the state is third-party cited, the owner may be PR, partnerships, reviews, directories, or analyst relations. A first-party page is not always the best answer source. For trust, comparison, and "best provider" prompts, AI systems may prefer sources that are not the vendor's own domain. The repair is to make those sources accurate and connected to the canonical page.

If the state is hallucinated or inaccurate, the owner is entity correction. Do not celebrate a brand mention that carries false information. Save the evidence, identify which source may have created the error, publish or update a correction source, and retest with the same prompt.

Method

The severity score

Use severity to prevent the team from treating every state as equally urgent.

Severity should combine three inputs: commercial value of the prompt, confidence that the state is repeatable, and risk of the current answer. One weird answer is evidence. Two or more repeated patterns across engines or retests become work.

Method

What "cited but not absorbed" means

"Cited but not absorbed" is the most overlooked state.

It means the answer linked to the page but did not actually use the page's distinctive claim, definition, data, process, or comparison. The citation exists, but the answer could have been written without the source.

This matters because a citation is not always a proof of influence. The answer may cite a page as a loose support link while drawing the actual explanation from another source or from generic model knowledge. Emerging AI visibility research is moving in this direction: the arXiv paper on per-entity bias mapping argues that aggregate mention and citation metrics are insufficient because entities have different error profiles ⁶.

In practice, save two fields for every cited answer:

• Citation selection: was the URL listed as a source?
• Answer absorption: did the answer use the page's definition, number, distinction, example, or workflow?

If selection is positive and absorption is weak, the page probably does not need more backlinks first. It needs stronger extractable answer units.

Method

How to use the taxonomy in a dashboard

The dashboard row should be small enough to run weekly.

If a dashboard has no action field, it is only a measurement archive. If it has no evidence field, it is a memory hazard. If it has no next retest date, it will quietly become stale.

Method

How ContentOS should use answer states

ContentOS should treat the answer state as the first field in a repair brief.

A source-pack repair brief should include:

• target prompt;
• engine, region, and language;
• observed answer state;
• answer text or screenshot;
• cited URLs and source types;
• desired canonical URL;
• target claim;
• missing evidence;
• required first-party page edits;
• required third-party/source graph work;
• internal links to add;
• FAQ/schema requirements;
• owner;
• retest date.

This keeps the writing task bounded. "Improve the article for AI Search" is vague. "The page is mentioned but not cited for three how-to prompts; add a direct-answer block, source markers, and related links to the canonical workflow page" is a task an agent and editor can complete.

It also protects quality. If the state is third-party cited, ContentOS should not automatically rewrite the first-party page. It should ask whether the third-party source is the correct citation surface for that prompt. If the state is hallucinated, ContentOS should not amplify the false claim in the page. It should create a correction plan.

Method

How often to retest states

Use the same cadence as the citation repair loop:

This cadence is not magic. It is a guardrail against two common mistakes: declaring victory from one lucky answer, or rewriting a page before it has had a fair discovery and retest window.

Method

What the taxonomy should not do

The taxonomy should not become a fake precision machine.

AI answers vary. Interfaces change. Google says there are no special additional requirements for inclusion in AI Overviews or AI Mode beyond the fundamentals of Search, and that site owners should continue to focus on useful, people-first content ⁵. That means the taxonomy is not a secret ranking formula.

Use it as an operating vocabulary. It helps teams talk about the same failure without mixing three different problems into one sentence. "We are invisible" is too broad. "We are mentioned but not cited on how-to prompts, and competitors are cited on comparison prompts" is actionable.