From AI visibility finding to 30-day work plan

The report arrived three weeks ago. Forty pages: where the AI platforms mention you, where they cite you, where they recommend the other guys. Everyone agreed it was eye-opening. Nothing has shipped since. The report is aging in a shared drive next to last year's brand audit, and the answers it measured are moving without you.

Reports pile up because a finding is not an instruction. “We lose the recommendation on mid-market questions” does not tell anyone what to publish, who approves it, or how you will know it worked. That routing is the plan, and it fits in thirty days if you keep the unit of work small: six failure types, one owner each, one correction each, one retest at the end.

Findings fail in six ways

In my work, findings keep reducing to the same six types. This is the routing table I use; the rare finding that fits none of them gets its own row and its own owner.

Two notes on the table. First, every row shares one acceptance test: the correction is live, machine-readable, and approved on the record by a named person before it ships. That named approval is the human gate. Second, the owner column is the point: a report stalls when every finding implicitly belongs to marketing. And the corrections are deliberately small. One good page beats a content sprint, and Google's AI-features guidance warns that scaled content variations produced to chase AI rankings violate its spam policies. The plan should never smell like that.

Grade before you route

Not every finding deserves a row in a 30-day plan. Grade each one on two axes: confidence and consequence. Confidence: did this appear once, or across platforms and repeated runs? A finding from one screenshot is a hypothesis. Consequence: is this a question buyers actually ask on the path to money, or a trivia query? The plan takes the high-high quadrant first, and it takes fewer findings than you want. Six to ten corrections in thirty days, each done properly, beats thirty started.

This grading is also where you say the uncomfortable thing out loud: some findings content cannot fix. If the models prefer competitors because third parties vouch for them and nobody vouches for you, the gap is earned coverage, not copy. If the honest comparative evidence favors the other product, the gap is the product. Route those out of the editorial plan and to the desk that owns them, visibly. A plan that pretends words fix everything burns a month proving it.

One finding, routed, with made-up details so the shape is visible. The report says two platforms describe you as a staffing agency; you build software. Type: category ambiguity, with a side of entity confusion. Owner: positioning and web. Smallest correction: a definition page that names your category plus Organization markup with consistent name, url, and sameAs links. Gate: your head of marketing approves the page before it goes live. Retest: the same locked category questions next cycle, watching whether the description moved. One row, one owner, one artifact, one check.

The weeks, concretely

Week one: route and lock. Grade the findings, pick the slate, assign owners, and lock the acceptance test per correction. Also lock the retest: the same questions in the same words, on the same platforms, from the same locked library the report used. If the report did not use locked queries, build the small library first; the query library canvas covers how.

Weeks two and three: ship the corrections. Each correction is a page, a markup change, a rewrite, or an outreach action, and each one passes a human gate before it goes live: someone accountable reads it and approves it. For entity corrections, the machine layer is the work: consistent Organization markup with sameAs links that unambiguously indicate identity, one canonical about page, the same name and url everywhere. For evidence corrections, follow the boring rules that survive every algorithm cycle: helpful, reliable, people-first content with crawlable links and structured data where it applies. The Ranqo preprint found roughly 78% of AI-answer citations pointing at corporate websites, a vendor figure to hold loosely, but directionally it says your own domain is where the evidence work pays.

Week four: retest and read. Rerun the locked queries. Check the free first-party slice too: Search Console's generative AI performance report covers Google's AI surfaces, where the report is available. Then write the honest reading: what moved, what did not, what shipped late, and what the next cycle takes on. Movement after one cycle is a signal. The trend across three cycles is a result.

Plans are suggestions with owners

One framing note so the plan survives contact with your organization. NIST's AI RMF Playbook describes itself as “neither a checklist nor set of steps to be followed in its entirety”: guidance to borrow from as your case demands. Hold your own plan the same way. The routing table is a model, and the six types will not cover every strange finding a platform produces. What must survive every adaptation: one owner per finding, one gate before anything ships, one locked retest at the end. Everything else bends.

How I run this

This loop is my method: query the platforms, decode why the answers prefer whoever they prefer, engineer the smallest honest correction, verify by rerunning the same locked questions. I run it on a production research engine I helped architect and build at Trinzik, the Austin AI company I co-founded; Trinzik operates it for clients. The engine supplies the findings and the retests; the routing table above is how those findings become work that ships. And the numbers stay honest: retests report directional movement, and I say so in the report, because the platforms rerank on their own schedule and a single cycle proves less than a dashboard wants you to believe.

If you already have a report gathering dust, the fastest unstick is mechanical: print it, mark every finding with one of the six types, and put a name next to each mark. The plan writes itself from there.

Questions worth asking next

Sources

1. Google Search Central, Search Essentials (helpful content, crawlable links, structured data). https://developers.google.com/search/docs/essentials
2. Schema.org, Organization type (name, url, sameAs, entity disambiguation). https://schema.org/Organization
3. NIST, AI RMF Playbook (suggested actions, explicitly not a checklist). https://airc.nist.gov/AI_RMF_Knowledge_Base/Playbook
4. Google Search Central, "AI features and your website" guidance (scaled-content warning). https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
5. Google Search Central Blog, "Introducing Search Generative AI performance reports in Search Console," June 2026. https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports
6. Pratyush Kumar (Ranqo), "Generative Engine Optimization at Scale," arXiv preprint 2606.20065, June 2026. https://arxiv.org/abs/2606.20065