When a brand is absent from AI answers, misrepresented in them, or losing share, the temptation is to prescribe content, because content is the lever everyone knows. But an AI answer is built through a pipeline, and a failure at each stage has a different fix. Prescribing content for an access-layer problem wastes a quarter. This chapter is the diagnostic: walk the pipeline in order, test each gate, and fix the first one that fails.
The five gates, in pipeline order: access (can the bots fetch you), retrieval (do you get found), selection (do you get read), citation (do you get credited), memory (does the model know you without searching). The first four belong to live retrieval and the fifth to model memory, the split from Model Memory and Live Retrieval; the pipeline stages behind gates two through four are walked in Retrieval and Grounding: How an AI Answer Gets Assembled.
One rule governs every test below: run it repeatedly before concluding anything. Answers vary across runs by design, so a single clean answer does not prove health and a single absence does not prove a problem; the measurement case for treating visibility as a distribution is empirical and published.Don't Measure Once: Measuring Visibility in AI Search, on why one-off observations mislead. The sampling protocol is in Measuring AI Search Visibility.
Gate 1: access
Test whether AI crawlers can fetch your pages at all. Check robots.txt for the relevant bot tokens, check whether your CDN or firewall blocks them, and fetch key pages as those user agents to see what actually comes back, including whether the content renders without JavaScript. The methods and the bot roster are in Technical Accessibility for AI Crawlers. For Google specifically, eligibility rides on ordinary indexing: AI Overviews and AI Mode retrieve from the core Search index through retrieval-augmented generation, so a page that is not indexed and snippet-eligible cannot appear, and Google's stated position is that normal SEO best practices are the access work.Google, Optimizing for AI features on Google Search: AI features are grounded in the Search index via RAG and query fan-out, and standard indexing and snippet controls govern eligibility. An access failure explains total absence across searched answers; if access checks out, move down the pipe.
Gate 2: retrieval
Test whether searches in your category pull your pages into the candidate set. Run your panel prompts with search enabled and record which domains the answers fetch and cite; then run the corresponding queries on the underlying engines (Google, Bing, Brave, depending on platform) and check where you rank, because an AI product retrieving through an engine where you rank nowhere will not find you. Retrieval failure looks like this: your brand absent from searched answers while the pages the answers do use come from queries you should own. The fix is classic ranking work plus query coverage, and it belongs to the chapters in the retrieval rows of the table in Model Memory and Live Retrieval.
Gate 3: selection
Being retrieved is not being read: a fetched page must survive chunking, scoring, and reranking to enter the model's context. The symptom of a selection failure is subtle, and log-based where you can see it: your pages get fetched by the answer-time bots, your rankings are adequate, and you still do not appear in the answer text. The usual causes are structural, passages that do not stand alone, answers buried mid-page, key claims spread across paragraphs, and the fixes are Content Architecture: Building for the Chunk and Writing for Retrieval and Citation.
Gate 4: citation
Being read is not being credited. Models paraphrase far more than they cite, and the gap is measured: Reddit is the most retrieved third-party domain in OppAlerts' data while being credited about once per 100 retrievals, and the largest credited bucket is brands' own sites at 53.9% of named citations.The retrieval-versus-citation analysis in the OppAlerts ranking-factors study. Method and tables in The Research Behind This Guide, part of the AI Search Visibility research. If your material shapes answers without your name attached, the citation-side levers, quotable claims, original numbers, being the primary source rather than the aggregator, are in Evidence, Expertise, and Original Research. Weigh this gate honestly: influence without credit is a smaller problem than absence, and for some programs it is not worth a quarter of work.
Gate 5: memory
Test what the model knows with search off: ask it directly about your brand, your category, and your competitors, repeatedly, through the API. Three failure classes come back. Unknown: the model has no association between you and your category, which is the slow-clock work of Entities: Becoming a Thing the Machine Knows and the fame signals in What Actually Correlates with AI Search Visibility. Outdated: the model describes your old pricing, product line, or ownership, which is the training-cutoff problem covered in Freshness, Cutoffs, and Temporal Accuracy, and retrieval is the only fast correction. Wrong entity: the model conflates you with another organization sharing your name, which shows up as descriptions that mix two companies' facts, or citations pointing at a domain that is not yours. Wrong-entity problems are disambiguation work: consistent naming, structured data, and entity records that separate you from the collision.
Symptoms to gates
| Symptom | Check first |
|---|---|
| Absent everywhere, searched and memory answers alike | Access, then memory |
| Present in memory answers, absent once search runs | Retrieval: competitors win the rankings underneath |
| Fetched by bots but never in the answer text | Selection |
| Ideas present, name absent | Citation |
| Named, but facts are stale | Memory: outdated; push current facts into retrieval |
| Named, but facts belong to someone else | Memory: wrong entity |
| Share falling on one platform only | That platform's model version and search engine, before your own site |
The 30-minute version, and the recurring one
The quick audit: ten category prompts on the two platforms your audience uses most, search on and off, three runs each; robots.txt and firewall check for the major bots; one direct memory question about your brand per model. That is enough to place you at a gate, and wrong enough often enough that the finding should trigger the full version, not a work order.
The recurring audit is this same diagnostic on a cadence, quarterly for most programs, monthly during active work, run against the time series from Building a Visibility Tracking System so changes are caught as they happen rather than reconstructed later. The deliverable is one page per audited category: gate status, what changed since last time, and the single highest gate with a failure, because the first failing gate is where the next quarter's work goes.