Free GEO Playbook: a step-by-step method for getting named in AI answers

Of 54 questions both engines answered, 36 shared not one URL. Work each engine separately. Only about 1% of AI Overview visits click a link (Pew, 2025, US sample): the prize is being named. Snapshot via model APIs, 23 Sep 2026.

Baseline: Google's organic top ten for the same question, same day, Singapore. By registered domain: AI Mode 53%, ChatGPT 25%. ChatGPT cited no forum or social pages in round one; asked whether one tuition centre was good, it cited a local parents' forum twice. Model APIs, 23 Sep 2026; the app may differ.

These are pass or fail: a firewall blocking AI, prices in images or mismatched addresses, and every page scores zero. Save answers before changing anything. Don't just append an allow block to robots.txt; it can cancel existing rules. Major AI crawlers skip JavaScript (Vercel & MERJ, 500M+ fetches).

Unsure? Go stricter: too loose can mean a fine, too strict only costs a few tactics. No licence but selling regulated products like supplements or devices? Your copy is still regulated. Strict-sector red lines: no "from" prices or ranges, no testimonials or before-and-after photos, no competitor comparisons, no asking for reviews, no paid Best or Top list spots.

Source questions from customers' words before buying, ad keywords, autocomplete and Search Console, and competitor pages. Never let AI invent them. Asking just "divorce lawyer", AI Mode listed 0 firms in 10 cities; adding "near me" restored every one (PRWeb, overseas). Our 6 Chinese questions drew 64 citations; fewer than 5 were Chinese pages, all in AI Mode.

On 15 of 18 law pages, the copied line opened a paragraph. Counter-example: an H2 opening "is a complicated matter" was skipped; a direct answer in the same page's FAQ was copied. Tables go whole, headers included. Positions are estimated by eye; answers were stored to 600 characters.

8 of 8 cited dental price blogs put the year in the title; 10 of 11 cited clinic pages used a statement H1. Exact numbers beat tidy layout: one tutor agency's card said $35–$55, its table $25–$35; AI used the table. Supporting data: price queries trigger AI Overviews 92%–97% of the time (Whitespark, 540 US queries).

These are counts on cited pages, not experiments: they aren't what gets a page cited, but that doesn't prove they're useless. Cited pages ran as short as 40 words; the median was about 1,300–1,500. On medical pages, do name the doctor. 88.2% of sites blocking GPTBot still get cited (BuzzStream).

Pool the sources across your 30 answers; the 20 most frequent are your targets. Correct what's wrong; if you're missing, start with free official directories and association profiles. Supporting data: 84% of AI citations come from others writing about you, 0.3% from paid placements (Muck Rack). Law and CRM rest on one question each.

Re-ask your questions three times in one week to see how much results swing on their own; only bigger moves are real. Switch questions or tools and you can't compare. Put a number only you have on your price page, then search next month's answers for it. Nothing moving? Recheck access and identity. Watch just one? Make it errors.

The full playbook, in English and Chinese: 17 chapters including 4 appendices, a blueprint for each of 46 page types, sources and sample sizes for key numbers, plus a Dental & Aesthetics Edition and a Law Firms Edition. The open-source version on GitHub is in English, ready for your AI assistant to check your site.

Want help with round one? admin@canlah.ai · WhatsApp +65 8110 0234