Open-source AI SEO
for

set up treg - https://treg.to/llms.txt then run the AEO workflow below. GOAL: 4-step audit of whether AI recommends the user's product. User decides at each ◆ gate. Step 5 (3-4 weeks later) measures movement. ═══ SETUP ═══ - `audit.md` with baseline appendix exists → ask: Step 5 | new audit | stop. - INPUT, one form (interactive card if available): product name | what it does (1 line) | product type noun | 2-3 seed keywords | brand | domain? | product URL? | price | where sold | features | outreach contact | country (US) | language (en). - Seeds must match the product type. Mismatch → flag and ask before spending. - PROGRESS at each step and gate: `[■■□□] Step 2/4 · <name> · task 1/2 · spent $0.06 / $0.30` + now/next. - GATES ◆: chart + 2-line reasoning → options, recommended first, always "stop". Before paid calls: About card (what · why · result · tools · calls · est. cost) → "Run Step X (~$cost)". Stop → build the report from what's done. RULES - treg: describe each endpoint first; quote its price. `dataforseo.x.*` takes an array body. - Python does all math. Every number the user sees comes from code output, re-checked before sending. CSV → `audit-step{1-4}.csv`. - minmax over the current set (0 if flat); cap outliers at p95 first. - URLs: strip `#…`/`?…`, dedupe, drop google.com/search links. - Brands: dictionary match, first-appearance order, links stripped. After each batch, add unmatched bold/linked brand names and re-run. - Fail → 1 fallback retry → else "unavailable", excluded from N. - Never invent specs, reviews or anecdotes. Output = drafts; the user publishes. - Labels: header "Step N of 4 · <name> · Check Gx" → finding as a plain-sentence title → "What these mean" legend (every column/axis/color) → "How X works" (formula in words + 1 real worked row). - Plain words for a small-business owner: DA → "site strength (0-100)", CPC → "what advertisers pay per click", intent → "ready to buy?". ═══ STEP 1 · What customers ask ═══ ~$0.04 1. `seranking.google.keywords.ideas` per seed (GET, source=cc, limit=40, sort=volume) → vol, CPC, difficulty. Merge, dedupe, drop off-topic rows (list them). 2. `dataforseo.x.dataforseo-labs-google-search-intent-live`, 1 call. I = transactional 1 | commercial 0.8 | informational 0.2 | navigational 0. BuyerScore = I(0.50 C + 0.35 V + 0.15 E); C = minmax(CPC), V = minmax(log10(1+vol)), E = 1 − diff/100. Keep I ≥ 0.8. 3. Merge look-alikes (reordered/plural) into groups, score = max. Keep the top 8. ◆ G1: table keyword | intent | CPC→C→×.50 | vol→V→×.35 | diff→E→×.15 | score, + 5 best cut ✗ and why big ones were cut. keep | edit | weights | stop 4. `treg.google.serp.organic` per group → People Also Ask. Keep choose/compare/buy questions, merge same-meaning ones. QuestionScore = max BuyerScore × (1 + 0.1(F−1)), F = groups sharing it. ≤10 questions. Rewrite as [situation] + [need] + "which should I buy?". No PAA → from keyword, tagged. ◆ G2: rank | question | source PAA | BuyerScore | F | score. approve | edit | swap | stop ═══ STEP 2 · How AI recommends ═══ ~$0.015/question Engines (parallel, 300s timeout): ChatGPT `cloro.ai-search.chatgpt.scrape` · Perplexity `treg.ai-search.perplexity.answer` · Gemini `cloro.ai-search.gemini.scrape` · AI Mode `cloro.google.serp.ai_mode` (gl/hl). Fallback: matching `dataforseo.x.ai-optimization-*` / `serp-google-ai-mode-live-advanced`. Per answer: m named, p position (m/p = 0 if m = 0), n brands, c own domain quoted. - VS = mean 100(0.5m + 0.3 m/p + 0.2c), also per engine · MentionRate = Σm/N · Share = Σm/Σn · CompScore_b = 100 Σ(m/p)/N - Sites/pages: Cite = Σ 1/source-position · QuotedIn = % of answers · Vis = 100 Cite/N · Infl = Vis/max · DomainScore = Cite(1 + 0.1(Q−1))(1 + 0.1(E−1)). Type each site. 1. Q1 on all engines. ◆ G3: per engine: quoted sites → brands (behind-the-scenes searches only if returned). all | ChatGPT + Perplexity | edit | stop 2. Full run. ◆ G4: tiles VS/MentionRate/Share/SiteVis_user · competitor bar · question × engine grid (✓/✗ + who is named 1st) · top-15 sites table · who wins, why. continue | re-run failed | stop ═══ STEP 3 · Websites AI reads ═══ ~$0.09 1. Candidates = normalized pages minus brand/retailer pages; flag pages about another product type. 2. `treg.web.url.metrics` (~$0.007) on the 12 most-quoted → PA, DA. Unmeasured → same-site PA, tagged. ◆ G5: scatter PA vs PageVis, color = type (top-left = easy wins) + table. continue | exclude types | stop 3. `treg.web.extract` top 15 → brands listed, criteria, date, contact route, covers product type? Opp = 0.40 Infl + 0.20 Gap + 0.15 (1 − PA/100) + 0.15 Access + 0.10 Fresh Gap 1/0.5/0 (absent + rival / both absent / present) · Access 1 contact-affiliate / 0.5 author / 0 · Fresh 1 ≤12mo / 0.5 ≤24mo / 0 ◆ G6: stacked Opp bars, top 10, one line each incl. fit. approve | swap | weights | stop ═══ STEP 4 · Get on the map ═══ ~$0.03 0. ◆ G7: confirm facts (from URL via `treg.web.extract`, else the form) and check question fit. Most don't fit → recommend redoing Step 1 with matched seeds. correct | fix | redo | stop Search with 2-5 word keyword forms of the questions. - Review/blog: Impact = Opp. Pitch ≤120 words. - Reddit `scrapecreators.reddit.search.posts` (relevance, year, trim; link = `url`): Impact = 0.4 minmax(log10(1+ups+2·comments)) + 0.3 exp(−days/180) + 0.2 Rel + 0.1 Cited → top 10 - YouTube `treg.youtube.search.videos`, title must contain the product noun, no competitor channels: Impact = 0.4 minmax(log10(1+views)) + 0.2 exp(−days/365) + 0.2 Rel + 0.2 Cited → top 5 channels - Own page: Impact = 1 − share of questions answered in the first 2 sentences (no page → 1). - Priority = Impact / Effort (1 reply · 2 pitch · 3 content). Thirds = High/Med/Low. Tag fit per item. ◆ G8: grouped list; per item `#1 · [channel] target (link)` · `Impact 0.88 ▓▓▓▓▓▓▓▓▓░ · Effort 2 · Priority 0.44 · fit` · why · next step. all | High only | drop | stop 5. Drafts: Reddit answers the question first, mentions the product only where it fits, and ends with a disclosure line. Plus pitches, YouTube outreach + 3 angles, and a page draft (direct answers, comparison table, specs, FAQ JSON-LD). ═══ FINISH ═══ - `audit-report.html`: one file, Chart.js (cdnjs), light/dark, tabs Summary · Steps 1-4 · Glossary, every external item linked. Summary: tiles, top 5 competitors/sites/targets/actions, spend vs budget. - `audit.md`: tables, drafts, baseline appendix (date, locale, brand, domain, facts, verbatim questions, endpoints + params, brand dictionary, raw answers). CSVs alongside. - Final message: 3 findings + next action + "Come back in 3-4 weeks and run this prompt again for Step 5." ═══ STEP 5 · Did AI start recommending you? ═══ ~$0.15 1. Load the baseline; change nothing (questions, engines, params, dictionary). 2. Ask each action's status: done | partial | not done. ◆ G9: show days since baseline (<21 → warn). run | wait | stop 3. Noise = mean per engine of (max − min VS) over 3 runs of Q1. |Δ| ≤ Noise → "could be chance". 4. Re-run everything, recompute Step 2 metrics, Δ for each. Gained/Lost = answers flipping named ↔ not named. NewCites = target pages newly quoted. Credit an action only if its target is newly quoted or linked. ◆ G10: "AI now recommends you in X of Y answers (was A of B)" · before → now cards with a noise tag · per AI · per question · leaderboard · site movement · what worked · next 3 actions. Append to the report and audit.md as the new baseline.