Most teams have never once measured whether AI assistants mention them - which means their AI visibility strategy, whatever the deck says, is a guess. This playbook replaces the guess in one working week: a fixed question set, systematic measurement across models, stage-by-stage diagnosis of every absence, and a ranked move list. It is the audit the AI Visibility agent runs as a mission; here it is as a method you can execute with any tooling.
What the audit answers
Three questions, in order of increasing value. Presence: for the questions our buyers actually ask, do we appear in the answers - cited, named, or both? Displacement: where we are absent, who is present - which rivals own which answers? Diagnosis: for each absence, which stage of the pipeline that produces AI citations dropped us - never retrieved, retrieved but not selected, or selected but not attributed? The third question is what turns the audit from a scoreboard into a work plan, because each stage has a different fix.
Step one: build the question set
The instrument is a fixed set of 20 to 40 questions phrased the way buyers phrase them - conversational, contextual, occasionally sloppy - never the way marketers write keywords. Source them from four places: sales-call questions (what prospects literally asked), Search Console query data for question-shaped terms, People-Also-Ask mining on your head terms, and the comparison questions your category forces ("X vs Y", "best tool for Z", "is X worth it"). Spread across funnel stages: definitional (top), comparative (middle), and decision-adjacent (bottom, where a mention is worth the most).
Step two: measure across models
Run every question against every assistant your buyers use, record the full answers, and extract: were we cited (linked), were we named, at what position among mentions, and who else appeared. Then - this is the discipline most audits skip - run the set again days later, because generated answers vary run to run and a single pass confuses variance for signal. The honest metric is presence rate across the set per model, per the method in share of voice, measured properly. Manual execution works at 20 questions; past that, tooling or a mission (the agent runs both passes and the extraction in one go) stops the audit from consuming the week it was meant to organize.
Step three: diagnose every absence
| Symptom | Stage | Fix family |
|---|---|---|
| No page of ours could plausibly answer; we rank nowhere near the topic | Retrieval | Technical citability + content that targets the question |
| Our relevant page ranks, but the answer quotes others | Selection | Retrofit: direct answers, liftable chunks, specificity |
| Our points appear paraphrased; our name does not | Synthesis | Quotable phrasing, entity consistency, corroboration |
Retrieval failures route to the AI search optimization checklist and, where the page simply does not exist, into the content plan. Selection failures route to passage-level retrofits per the LLM SEO formats article. Synthesis failures are the slow burn - entity and corroboration work from the generative engine optimization guide's third layer.
Step four: rank the moves
Score each absent question on two axes: intent value (a decision-adjacent question outweighs a definitional one) and fixability (a selection retrofit on an existing page ships this week; net-new topical authority takes a quarter). Work the high-value, high-fixability quadrant first - it usually contains a handful of retrofits that move the presence rate visibly within weeks, which buys the program credibility for the slower work. Technical blockers, if step three found any, jump the queue entirely: they are cheap, fast, and gate everything else.
Step five: set the cadence
The audit's real product is the baseline; the value compounds on the re-run. Weekly scorecard on the frozen set (same models, presence rate per model, deltas highlighted), monthly technical re-audit, and the move list reranked as items ship. As a scheduled mission this costs minutes a week and posts its delta to the channel where your team will actually see it - the AI Visibility playbook shelf has the runnable versions.
The presence-rate calculator
Your headline number, computed from any measurement pass:
The working checklist
Frequently asked questions
What is an AI visibility audit?
A systematic measurement of whether AI assistants cite or name your brand on the questions your buyers actually ask - across models, twice for variance - with each absence diagnosed to a pipeline stage and converted into a ranked fix list.
How many questions should the set contain?
20 to 40, sourced from real buyer language across funnel stages, then frozen. Fewer lacks coverage; more makes manual measurement collapse and adds little signal. The set is a longitudinal instrument - stability is its value.
Why measure twice?
Generated answers vary run to run; a single pass confuses variance for signal. Two passes days apart, aggregated into a presence rate, is the minimum honest measurement - scheduled weekly runs are better.
What is a good presence rate?
The useful benchmark is relative: your rate versus the rivals appearing in your absences, on the same set. Early audits commonly find single-digit rates against established competitors - which is the point of running one.
How fast can the rate move?
Technical fixes and selection retrofits on existing pages can move measured presence within weeks, because engines refresh continuously. Entity and corroboration gains build over months. The audit tells you which kind of work you need.
Sources
- GEO: Generative Engine Optimization (Aggarwal et al., 2023)
- Google Search Central - fundamentals that govern the retrieval stage
Every playbook on this blog ships as a runnable mission.
Open a workspace and the playbook library is waiting - describe the outcome and the agents carry it end to end, on your plan's monthly credits.