How Do I Run a 90-Day GEO Pilot?
Run a 90-day GEO pilot in three phases: baseline (weeks 1-2), build (weeks 3-10), and decide (weeks 11-13). Scope it to one product line and a fixed set of 20-30 buyer prompts, measure every engine before you touch anything, ship 15-25 answer-first pages plus technical fixes, then make a go/no-go call on measured lift — not vibes.
Why scope narrow before you scale?
A pilot exists to prove the mechanism works for your category, not to blanket the site. Pick one product line, service, or persona so the prompt set stays coherent and the signal is readable. Broad pilots dilute the result: if you optimize forty topics at once and share of voice ticks up, you cannot attribute the gain. Baseline the narrow slice, improve it, and the lift is unambiguous.
What do you lock in the first two weeks?
Freeze three things so the endline comparison is honest:
| Artifact | What it fixes | Why it matters |
|---|---|---|
| Prompt set (20-30) | The exact buyer questions you track | Same prompts run at baseline and endline |
| Engine list | Which platforms count | Answers differ per engine; measure all in scope |
| Success bar | Numeric go/no-go threshold | Prevents post-hoc rationalization |
Run each prompt in a fresh, logged-out session to strip personalization, and record who gets mentioned and cited. This baseline is the control you measure the entire pilot against.
What ships during the build phase?
Weeks 3-10 produce two workstreams. First, content: answer-first pages for the highest-intent prompts, structured with question-shaped headings and 40-80 word self-contained passages the model can lift. The GEO research by Aggarwal et al. (KDD 2024) measured 30-40% higher generative visibility for content adding statistics, quotations, and citations. Second, technical: confirm AI crawlers like OAI-SearchBot and PerplexityBot are not blocked in robots.txt, add or fix schema, and submit new URLs via IndexNow so Bing (ChatGPT Search's index) picks them up in weeks, not months.
How do you decide at day 90?
Re-run the identical prompt set and diff it against baseline. Judge three signals: mention-rate change, share-of-voice change, and any new first-position citations. Because retrieval-based mentions can surface in 2-8 weeks but parametric knowledge lags training cycles, weight retrieval engines more heavily in a 90-day window. If you clear the bar you set on day one, scale the playbook to the next product line; if not, diagnose whether the gap was crawlability, corroboration, or content quality before spending more. Tools like Menra automate the repeat measurement so the endline is one export, not a week of manual prompting.
Frequently asked questions
- How many pages should a GEO pilot produce?
- Enough to move a scoped prompt set — usually 15 to 25 answer-first pages plus technical fixes on existing high-intent pages. A pilot proves the mechanism, not scale, so resist publishing hundreds of thin pages before you know the model retrieves them.
- What is a pass/fail bar for a GEO pilot?
- Set it before you start. A common go decision is a measurable rise in mention rate or share of voice on your tracked prompts across at least two engines, plus at least one first-position citation you did not hold at baseline.
Keep exploring
See how AI engines talk about your brand — track mentions across ChatGPT, Perplexity, Claude, Gemini and 5 more. Start with Menra