AI Tab — Data Integration Plan

Reuses the tab as-is. Backend-only for the first item; small UI for the second.

• Real trend line via historical/live. Replace the sparse self-built mentionsHistory with a true monthly series (mentions + AI search volume) back to Aug 2025. A brand-new customer sees a full chart on day one instead of waiting weeks. One call, ~$0.10, per refresh.
• Keep the mentions trend as the Google+ChatGPT layer. The mentions number stays honest and deep for the 2 engines the API actually has data for. It is not the "all engines" view — that comes from Phase 2's live checks (below). Label it clearly so nobody reads a "2 of 4" number as the whole picture.
• Verify endpoint naming. Our code calls search / top_domains / cross_aggregated_metrics; the current published API uses search_mentions / top_mentioned_domains / target_metrics. Confirm ours are still-supported aliases or migrate — cheap insurance against a silent all-zero tab.

New sections. This is where the tab gets its "wow" — real answers from the models, and real demand numbers.

• The all-4-engines view — "Here's what every AI says about you" (LLM Responses). This is the primary engine display. Take the customer's top questions and ask ChatGPT, Claude, Gemini, and Perplexity live. Show one row per engine: the real answer + whether the site is cited. This is the only path that can show all four uniformly — Gemini and Perplexity exist nowhere else in the API. Framed positively — "cited by 3 of 4 assistants," never a rank. Strongest demo moment and a genuine differentiator.
  Cost guardrail: token-priced and up to 120s each. Cap questions per site, use cheap models (Haiku / gpt-4o-mini / gemini-flash / sonar), run via the async Standard task flow (charged ~1¢ upfront, refunded) with postback_url, and cache — don't re-ask every page load. Perplexity is Live-only.
• "How much are people asking AI about this?" (AI Keyword Data). Feed the site's tracked keywords in, show AI search volume + its monthly trend. Reframes the tab from "who mentions you" to "here's the demand, and here's your slice of it." Cheap, batchable up to 1000 keywords/call.
• "Which of your pages AI quotes" (top_mentioned_pages/live). Move from domain-level to page-level citations — lets us say "your guide page is what AI pulls from" and target content accordingly.

Nice-to-have richness once Phases 1–2 land.

• Real consumer-UI snapshots (LLM Scraper). Scrape what ChatGPT / Gemini actually render for the customer's key queries — structured sources, brand entities, product/table blocks. ~$0.004/call. Great for screenshots and "this is the literal answer" evidence.
• Brand association tracking (top_mentioned_brands / brand_categories). Which brands AI ties to the niche — competitive-landscape view.
• Momentum widgets (timeseries_delta / timeseries_new_lost). "You gained N new AI mentions this month" — positive, movement-only framing that fits the product rule.

• Fan-out queries → content targets. Every LLM response already returns fan_out_queries — the sub-questions the model expands a query into. These are the literal prompts to write content for. Turn them into new_content plan items. Pure GEO signal, zero extra API cost.
• Citation-source gap. top_mentioned_domains / top_mentioned_pages shows which sources AI pulls from in the niche (Reddit, Wikipedia, directories, specific competitors). Reframe from a stat into a tactic: "AI trusts these places — here's where to get you listed or mentioned."
• Brand-fact monitoring. brand_entities + the live answers reveal what AI actually says about the customer — attributes, category, facts. Flag when AI is wrong (wrong pricing, "permanently closed", stale info). Correcting AI hallucinations about your own brand is one of the biggest GEO wins and nobody else surfaces it.

• Question coverage tracker. A fixed set of "money questions" per niche, re-run across all 4 engines on a schedule → citation status + trend per question per engine. The core GEO KPI: "cited for 6 of 10 key questions, up from 4." Positive, movement-only framing.
• Share of AI voice. target_metrics/live gives your mentions vs the niche total → "you appear in ~12% of AI answers here." Never a rank — a share that grows.

• GEO-readiness checklist. Schema.org / JSON-LD (we already do some), FAQ markup, clean extractable content, llms.txt — the technical factors that make a page easy for an AI to parse and cite. Surface as a plain-language checklist Scaup works through; we could even auto-generate and maintain llms.txt.

• Before / after citation. After Scaup publishes a fix, re-run the live query later: "After we added this FAQ, ChatGPT now cites you for X." Closes the loop between what we did and what changed in the AI answers — the most convincing proof of value we can show.