See, Fix, Prove: The Only AI Visibility Workflow That Survives Review

Most AI visibility programmes do not fail because the work was wrong. They fail at the quarterly review, when someone asks what the investment produced and the honest answer is a screenshot of a score.

This is a workflow problem, not a tooling shortage. The teams that keep their budget run a specific loop: see what the assistants say, fix what is causing it, and prove the fix moved the answer. Each step is only useful if the next one is possible.

Why the score is not the deliverable

A dashboard reporting "your AI visibility score is 34" answers no question anyone senior is asking. It is not comparable to anything, it does not indicate whether that is good for your category, and it does not suggest an action.

The question a marketing manager is actually asked is narrower and harder: did the thing we spent money on change what buyers see? Answering that requires knowing what buyers saw before, which means measurement has to predate the work. This is the single most common failure in the whole discipline. Teams do good work for two quarters, then try to demonstrate impact retroactively, and discover they have no baseline to compare against.

So the loop begins before the work does.

See: measure what buyers actually encounter

Seeing means knowing what assistants say when a buyer asks a question in your category. Three things determine whether that measurement is worth anything.

The questions must be unbranded. If you ask an assistant about your company by name, it will describe your company. That measures recall, not discovery, and it will make you feel better than the situation warrants. Real measurement uses the questions buyers ask before they know you exist.

Coverage must span engines. Brands routinely appear in one assistant and are absent from another. Those are different problems: a single-engine gap usually points to freshness or retrieval, while a flat gap across all engines points to how your category membership is described across the open web. Averaging them into one number hides the diagnosis.

Sampling must repeat. AI answers vary between runs. One check is an anecdote, and building a programme on anecdotes means every result is arguable.

The practical output of a good see step is not a number but a list: here are twenty buyer questions, here are the four where we are named, here are the sixteen where a competitor is named instead. That artefact is legible to anyone.

Fix: know specifically what to change

Seeing that you are absent is diagnosis without treatment. The fix step converts it into work someone can do on Monday.

In practice the causes cluster into a small number of patterns: no consistent description of what your company does, absence from the third-party pages models draw shortlists from, content written for branded rather than buyer questions, pages an engine cannot extract a clean claim from, and crawler access problems nobody has checked. Each has a different owner and a different timeline. The full diagnostic sequence is in why competitors appear in AI search but you don't.

What makes this step work is specificity. "Improve your content" is not a fix. "These six questions are answered by a competitor's comparison page that does not mention you, and your equivalent page is client-rendered so crawlers see an empty shell" is a fix, because it names the page, the cause and the owner.

Prove: connect the change to the movement

This is the step almost nobody does, and it is the one that determines whether the programme is funded next year.

Be honest about the ceiling first. Clean causal attribution is not available in AI search. Answers vary, engines update on schedules you do not control, and the third-party sources that shifted are usually not yours. Any vendor claiming precise causal attribution is overstating what the medium allows, and that overclaim is the fastest way to lose credibility with a sceptical CFO.

What is achievable is a defensible timeline: here are the questions we tracked, here is the baseline, here is what we changed and when, here is the movement afterwards, and here is the sampling that shows the movement exceeds normal variance. That is correlation stated honestly, and it is considerably more persuasive than fabricated causation because it survives scrutiny.

The requirement this places on tooling is specific: the history must be continuous and recorded on the same basis as the baseline. You cannot reconstruct it later.

Why splitting the loop across tools breaks it

The common setup is measurement in one tool, content work in another, crawler checks in a third, and reporting assembled by hand in a slide deck.

Each tool works. The loop still fails, and it fails at the handoffs. The baseline lives somewhere the content team does not look. The content changes are not timestamped against the measurement. The reporting is a manual reconstruction that happens once, before the review, and then never again because it took a day.

The practical consequence is that prove becomes optional, and once prove is optional the programme is judged on vibes. Teams that keep funding are the ones where the timeline assembles itself, because the measurement, the change log and the reporting share a spine.

This is the argument for a unified platform, and it is worth stating in its weakest honest form rather than its strongest marketing form: a single tool is not inherently better at any individual step, but it is dramatically better at the handoffs, and the handoffs are where this specific loop dies.

Running the loop

A workable cadence for a marketing team:

1. Pick the questions. Ten to twenty unbranded buyer questions with genuine commercial weight. Not "what is [category]" but the questions asked two weeks before a purchase decision.
2. Record the baseline across assistants before changing anything.
3. Diagnose which pattern is causing absence for the questions that matter most.
4. Fix in priority order, timestamping what changed.
5. Re-measure on a fixed interval, not when you feel like checking.
6. Report the timeline, not the score.

Step two is the one teams skip and cannot recover from later.

Where to start

Get the baseline today, before any of the work. The free AI Visibility Scorecard runs unbranded category questions across the major assistants and produces the list of where you are named and where competitors are named instead.

From there, the AI visibility tracker handles continuous measurement, AI Ranking organises it around buyer questions rather than keywords, and Visibility Trends is where the timeline that answers the quarterly-review question lives. If you are still choosing tooling, how to choose an AI visibility platform covers the evaluation framework in full.