Build your own AI Brand Intelligence tool

1. Main prompt paste first

Build me a Console app called "AI Brand Intelligence". It audits what AI assistants (ChatGPT, Perplexity, Google AI Overviews, Gemini, Copilot…) say about a company, compares that with what the company's own website says, and tells me exactly what to fix. Data comes from Ahrefs Brand Radar (existing AI answers; don't buy new prompt tracking). Store everything in Postgres, run long steps as background jobs, and keep every project separate.

WORKFLOW
1. Setup: I enter a domain. Auto-fill the brand name, aliases and likely competitors; I can edit them.
2. Discovery: read the homepage plus up to ~50 of the most relevant pages (services, about, pricing, contact, footer). If the live site is blocked, fall back to Ahrefs' crawl archive. Extract facts (founding year, location, team, prices, ratings, credentials, service list…). Every fact keeps its source URL and an exact quote, and you must check the quote word-for-word against the page text. Drop facts that fail the check.
3. Review: I approve, edit or reject facts and services. ONLY approved facts are ground truth. I can pick my top 3 priority services.
4. Analysis: pull Brand Radar answers that mention the brand, plus unbranded answers about each service. Extract every factual claim AI made about the company and check it against the approved facts.
5. Results: a 0–100 AI Brand Health Score, plus tabs for Overview, Action plan, Services, Brand facts, Website, Data & method. Clicking any finding opens the actual AI answers behind it in a side panel.

EVIDENCE RULES (non-negotiable)
- Every finding and every action must link to the AI answers (observation IDs) it rests on. No evidence, no finding.
- An AI not mentioning something is missing data, never an error.
- Compare like with like: "190+" and "over 200" are compatible; "$565–$700" overlaps "$565"; "starting at $189" conflicts with a $565 minimum. Check qualifiers, ranges, units and scope before calling something wrong.
- Changeable facts (prices, counts): an older AI answer is "incorrect" only if the site showed the approved value at that time (check an archived copy). Otherwise label it "needs verification" and leave it out of the score.
- Exclude claims about a different company with the same or a similar name, and never attribute third-party claims (review sites, directories) to the company without first-party evidence.
- Text missing from the pages you stored doesn't prove the site lacks it. Check every relevant page, the footer included.
- Keep sources apart: the company's own pages, Brand Radar answers, your own tests, and third-party pages AI cited. Record model, date, question and sample size.

SERVICES: THREE SEPARATE MEASURES
1. Knows you offer it: only explicit "Does [brand] offer [service]?" answers. Under 3 answers shows "Not tested", never a penalty.
2. Recommends you: unbranded "best provider for [service]" answers. Show who AI recommended instead, but never list AI platforms or job titles as competitors. Answers that name no provider are informational, not a loss.
3. Mentioned unprompted: descriptive only, never scored.
Sample labels: under 5 insufficient (no rating), 5–9 provisional, 10–19 limited, 20+ adequate.

SCORE
AI Brand Health = factual accuracy 35% + service recognition 25% + recommendation strength 25% + source readiness 15%. Any component without enough evidence is left out and the rest re-weighted, and the score is labelled "Provisional". Missing data never scores zero.

ACTION PLAN (up to 8 actions, ranked)
- Do the work, don't assign it: each action states the conclusion and the concrete change, the target page and section, why it matters, suggested copy, and how to check that it worked. Never make "review / research / investigate" the main step.
- Readiness labels: "Ready to implement" (a verified wrong or missing fact on a named page), "Recommended experiment" (anything aimed at getting recommended), or "Needs confirmation" (exactly one defined check plus the missing information).
- Suggested copy uses only approved facts or quotes from the site: no new figures, no superlatives, unknowns as [placeholders]. Never promise that a change will make AI recommend the company.
- Rank by business importance (top-3 services first), severity, frequency/recency, evidence strength, confidence and effort.

COST CONTROL
- Brand Radar reads and rebuilding from saved data are free to re-run.
- Optional "standardized tests" (asking AI models the direct and unbranded service questions several times each, through APIs with web search) must show the exact number of calls and the estimated cost, and run ONLY after I click approve.

DESIGN
Use the standard Console look. Show the score prominently with its four component bars, a 2–3 sentence plain-English diagnosis, and "what's working / biggest problem / biggest opportunity". Every rebuild gets a snapshot ID that the header, score, actions and exports all share.

Plan first. Tell me your plan and any decisions I need to make, then build it step by step and test it on my domain: [YOUR-DOMAIN.com]

2. Follow-up prompts (one at a time, after the core works)

Cause tracing

For every claim AI got wrong, trace where the wrong value came from: search the company's own pages and the pages AI cited for that value. Label the cause (the company's own site contradicts itself, a third-party page, AI mixed up two of the company's facts, confused with another company, hard to find on the site, or not stated on the site) and show the source URLs.

Rechecks

Add "Recheck this issue" to each action: once I've made the change, re-ask the same questions (estimate the cost first, run only on approval). Keep recheck answers separate from the baseline score and report observed change, not proven effect. Add an opt-in "Update analysis with rechecks" that adds each eligible answer once and keeps the old baseline score for comparison.

PDF + CSV exports

Add a client-ready PDF report (6–7 pages: score, diagnosis, wrong claims with sources, services, action plan, method) and CSV exports. Both use saved data only, make no AI calls, and include the snapshot ID, sample sizes and source links.

Public share page

Add a "Publish public web report" option. Freeze the current analysis (plus the AI answers it references) into a shared table, and render it on my public site at /ai-brand-report/<slug>, with no AI or Ahrefs calls per visitor. Design: slick dark mode only (color-scheme: dark, near-black navy background, orange for the brand name, score and primary highlights, blue for evidence, links and navigation, a distinctive display font, an animated score ring, green/red/amber result chips). Sections: hero with score and diagnosis, what AI gets wrong (with "Read the AI answers" pop-ups), knows-you vs recommends-you per service, website findings, the action plan with copy buttons, and the method. Add a dark 1200×630 share image for LinkedIn/X. Include Publish / Update / Unpublish in the app.

Tips

• Use Build mode, and let it plan before it codes.
• Test on a company you know well, so you can tell when the AI is wrong.
• If results look off, ask it to "show me the evidence behind this finding". Every claim should trace back to an AI answer and a page quote.