The 7 layers of AI visibility

"Are we visible in AI?" is seven different questions in a trench coat. Each fails independently, each has a different remediation, each maps to a different team or system. The Combot framework names seven measurable stages — Trained-on, Retrieved, Fetched, Cited, Sentiment & Accuracy, Recommended, Acted-on — and lines them up against the three Knowledge Modes underneath.

The 7-layer funnel

Trained-on → Retrieved → Fetched → Cited → Sentiment & Accuracy → Recommended → Acted-on

Each layer is independently measurable. Each fails differently. Each has a different remediation playbook. Together they map cleanly to a dashboard, to a board deck, and to the three Knowledge Modes (Memory · Search · Fetch).

Why seven, not five, not thirteen

Three things informed the cut:

1. A stage has to be a stage. The user's prompt is the input to the funnel. Synthesis is the output event, not an intervention surface. Community reputation feeds training data — it is not a downstream stage. We keep stages, drop inputs and outputs.
2. If two layers share a remediation, they share a layer. Training-data presence and entity-graph consistency are both fixed by authoritative public profiles, Wikipedia, Wikidata, schema.org. They become Trained-on. Sentiment and recommendation are different questions but both live in the answer text; they get their own layers, but accuracy travels with sentiment because the remediation overlaps.
3. The industry vocabulary is 4-7 concepts. Profound publishes ~6, Otterly 11 KPIs under a 3-step lifecycle, Ahrefs 4, Surfer 5, BrightEdge / Conductor / seoClarity / Brandsight 3-6, and academic papers cluster at 4-5 dimensions. A 7-layer funnel sits inside the conversation; it earns its precision instead of trading on novelty.

Layer 1 — Trained-on · Memory mode

What the model learned during pretraining, and how cleanly your entity is disambiguated inside that training corpus. Remediation: authoritative public presence, Wikipedia and Wikidata, schema.org sameAs, consistent NAP across the open knowledge graph, sustained community signal in the open corpora that feed pretraining.

Measure: tools-off probes per model (web_search and web_fetch disabled), declarative top-3 prompts, binary inclusion scoring. See Measuring Memory.

Move it: see Optimising Memory.

Layer 2 — Retrieved · Search mode

Whether your URL survives the retrieval step when the model invokes a search tool. The AI rewrites the user prompt into one or more queries, hits a search backend (Brave for Anthropic via MCP, Google for Gemini, undisclosed for OpenAI), and your page either makes the candidate set or does not.

Measure: cited-URL extraction per vendor, share-of-citations, search-ratio (Search vs Memory) per prompt. See Measuring Search.

Move it: see Optimising Search.

Layer 3 — Fetched · Fetch mode

Whether the AI can actually read the URL once retrieved (or once a user pastes it). Most AI fetchers do not execute JavaScript reliably. Client-side-rendered pages are invisible to the majority of the AI crawler population.

Measure: per-vendor URL-recall probes, fetch-readiness scoring, server-log analysis. See Measuring Fetch and Server logs for the AI era.

Move it: see Optimising Fetch and Lean Render; PrerenderProxy productises the playbook.

Layer 4 — Cited

Your URL appears as a source in the final AI answer — not merely "retrieved" but "kept" through the synthesis step. Citations land in vendor-specific structured fields: Anthropic's content[].citations[], OpenAI's message.content[].annotations[] with type: "url_citation", Gemini's groundingMetadata.groundingChunks[].

Measure: cited-URL share, citation prevalence across repeated samples (citation behaviour is non-deterministic; cite the academic uncertainty literature), per-vendor citation format parsing.

Move it: snippet engineering, structured data, llms.txt — covered in Optimising Search and Source mapping.

Layer 5 — Sentiment & Accuracy

How the answer frames your brand: positive, neutral, or negative, with what attributes attached, and whether the underlying facts are correct. Sentiment without accuracy is vanity; accuracy without sentiment misses competitive dynamics. We measure them together because the remediation overlaps — both require authoritative source presence and active fact-checking against AI outputs.

Measure: answer-text NLP for sentiment, fact-accuracy diffing against a verified knowledge base. See Fact-checking the machines.

Layer 6 — Recommended

You are not just mentioned — you are chosen. The model puts you in the top three when asked to recommend. This is the layer that converts to commercial outcomes and the layer that most directly maps to our north-star metric, AI Share of Recommendation.

Measure: per-prompt top-3 inclusion, weighted SOR (prompt weight × model weight × channel weight × recommendation), per-model variance.

Layer 7 — Acted-on

What happens after the recommendation. Did the user click through, ask a follow-up, complete a transaction? In agentic flows (booking, purchase, support routing), did the agent actually invoke the action? This is the layer that closes the loop with revenue.

Measure: referral attribution, agent-trace logs, downstream conversion metrics — the same data that classical analytics already collects, re-tagged by the AI surface that drove the visit.

How the 7 layers map to the 3 Knowledge Modes

How the 7 align with the industry

A 2026 survey of public AI-visibility frameworks shows the same compact vocabulary recurring across vendors and academic papers — presence, citation, share/position, sentiment, source ecosystem, business impact. The 7-layer model lines up against each:

What sits outside the funnel

Three categories of detail are real but not stages — they belong elsewhere in the playbook:

• Prompt intent is the input to the funnel. It lives in the prompt-portfolio work — brand recall, category discovery, comparison, trust, price, action — not in the funnel itself.
• Synthesis happens inside the model; it is the event that produces Layers 4-6, not a separate intervention surface.
• Per-user personalisation, vertical AI surfaces (Shopping, News), and feed/database integrations are per-vendor variants that show up inside Layers 2-4 rather than as standalone stages.

For teams that want engineering-level granularity, the underlying diagnostic breakdown — failure modes, sub-checks per stage, vendor-specific quirks — lives in the knowledge base. The lead funnel, the dashboards, and the executive story stay at seven.

Companions: AI Knowledge Modes · Tools for Modes · AI Share of Recommendation

Series: Optimising Memory · Optimising Search · Optimising Fetch · Measuring Memory · Measuring Search · Measuring Fetch