A First-Month Plan for an AI Search Programme
Spend the first month building the instrument, not chasing results. Week one: check which first-party AI data your properties already have (Search Console's generative AI report and Bing Webmaster Tools' AI Performance, as of September 2026), confirm your key pages are crawlable, and name owners. Week two: write and freeze a versioned prompt set, map each prompt to a page, and register exactly which surfaces you'll sample. Week three: run baseline window A with repeated runs and store every row. Week four: run window B a week after A, send a first report with a methodology note, and design one experiment with a control set and stopping rules. Ship no changes to the control topics. A lift reported in month one can't be separated from noise, because you don't yet know your normal variation.
The pressure in week one is to do something visible: rewrite a page, buy a tool, post a screenshot. Resist it. Everything you do before the baseline exists becomes impossible to evaluate later, and there's no way to take a 'before' measurement after the fact.
The four weeks
| Week | Ship | Skip on purpose |
|---|---|---|
| 1. Sources and reachability | Check which first-party AI reports each property shows; confirm key pages are crawlable, indexed and readable without JavaScript; name owners | Content changes and tool purchases |
| 2. The instrument | Prompt set v1, versioned and brand-free; prompt-to-page map; surface register; counting rule in writing | Running prompts before the set is frozen |
| 3. Baseline window A | Repeated runs on every prompt and surface, every row stored; first platform-change log entries | Interpreting a single window |
| 4. Window B and first report | Window B a week after A; report with a methodology note; one experiment written down with control set and stopping rules | Claiming a lift |
The month-one checklist
[ ] Search Console generative AI report: available / not available, per property
[ ] Bing Webmaster Tools AI Performance: site verified, data present?
[ ] Key pages crawlable, indexed, content present without JavaScript
[ ] Owners named: prompt set, sampling, pages, reporting
[ ] Prompt set v1 frozen: __ prompts, none naming the brand
[ ] Prompt-to-page map filled
[ ] Surface register filled and dated
[ ] Window A sampled: __ runs per prompt per surface, rows stored
[ ] Window B sampled a week later, rows stored
[ ] First report sent, with methodology note and a 'not measured' line
[ ] One experiment written down: treatment, control, dates, stopping rulesWeek one starts with data you may already have
Google's generative AI performance report in Search Console reached all websites worldwide on August 31, 2026, according to Google's announcement, and Microsoft launched an AI Performance report in Bing Webmaster Tools as a public preview in February 2026. Check both before designing anything. They cover Google's AI features and Microsoft's AI surfaces only, and they count appearances and citations rather than what the answer said, but they're free, they're first-party, and they tell you in an afternoon whether any of your pages are already showing up.
Why two windows, and why that still isn't enough
Two windows a week apart show you whether the numbers wobble, which is the first thing a stakeholder will ask. They don't tell you how much wobble is normal. That needs three or four quiet periods, which is why the first experiment's stopping date should sit well beyond month one, and why the first report says 'baseline' rather than 'trend'.
What to tell stakeholders on day one
That month one produces an instrument and a baseline, not results; that the first review with any chance of a signal is a set date, usually a quarter out; and that the report will include what wasn't measured. Saying this upfront costs one uncomfortable meeting. Not saying it costs every meeting after the first flat month.
Being straight about our own position
Fenn runs checks across ChatGPT, Claude, Gemini, Perplexity and Grok, and we'd like you to use it. Month one works fine with a spreadsheet and a fixed prompt list, though. What matters in these four weeks is the discipline of freezing the instrument before touching anything, and no tool supplies that for you.
FAQ
Can we skip the baseline if we're in a hurry? — You can, and every later report will open with an apology. A baseline can't be taken retroactively.
How many prompts in month one? — Enough to cover the buying questions that matter, few enough to run repeatedly on every surface. A smaller set run properly beats a larger one run once.
What if we already shipped changes before starting? — Then this period can't be attributed, and the honest move is to say so and start the baseline now. Don't construct a retrospective 'before' from memory.
Put this to work on your own website.
Fenn finds what your customers ask, drafts the articles and site fixes, and measures what ChatGPT, Claude, Gemini, Perplexity and Grok say about you — with every change waiting for your approval.