The grid that shows where AI search will cite you

AI search doesn't rank one page for one keyword. It lifts answers across related questions — and nobody can tell you, offhand, which of your pages actually answers which of your buyers' questions. The page-vs-question grid is the artifact that answers it, and it has always been too expensive to build: thousands of cells, each one a careful relevance judgement.

Until the judgement costs a fraction of a cent. Score one page against one question, repeat across the grid, and read the coverage like a heatmap. Strong cells are your citation surface. Empty rows are questions you have no page for — and right now, someone else's page is getting cited instead.

The architecture

The join is the expensive-looking part and it is just code. A hundred questions times fifty pages is five thousand cells — forty batch calls, cents of metered spend, minutes of wall clock.

Build it

curl $WF_URL/predict/batch \
  -H "authorization: Bearer $WF_KEY" -H 'content-type: application/json' -d '{
  "states": [{"buyer_question": "How do I find orphan pages?",
              "page_url": "/orphans",
              "page_excerpt": "Run a crawl, filter zero-inlink pages..."}],
  "policy": "seo_answer"
}'

rows = run("answer", cells)  # examples/seo_bulk.py
gaps = [r for r in rows if r["answer_relevance"] < 1.0]

Read the grid in three passes: defend the 3s (they earn citations today), improve the 1s and 2s (one section away from best-in-class), and build for the empty rows. That third list, sorted by question frequency, is the only content roadmap you need this quarter.

Where questions come from

The grid is only as good as its rows. Mine support tickets, sales call transcripts, community forums, and "people also ask" boxes — real buyer language, not keyword-tool paraphrases. Twenty real questions beat two hundred synthetic ones.

Pitfalls

• Excerpts, not whole pages. The first screen of content is what gets cited; score that.
• Confidence cuts both ways. A 2.5 with low confidence is a "probably answered" — spot-check before you celebrate.
• Rebuild quarterly. AI answer engines shift what "answered" means. The grid is a living artifact, not a report.