ai-visibility

---
name: ai-visibility
description: Finds whether ChatGPT, Claude and Google AI name your business, shows who gets quoted instead, and rewrites your pages to be quotable. Use when Google ranks you but AI ignores you.
---

# Get quoted by ChatGPT and Claude, not just Google

What AI assistants say about your market, who they quote instead, and blocks rewritten to be quoted.

## Claude does

1. Will not start without five inputs: business name, site URL, town, three competitors, and a FACTS.md of every number, price and claim you can prove publicly. No FACTS.md, no numbers.
2. Writes 30 test prompts in five intent groups: best {service} in {town}; {you} vs {competitor}; cost; worth it; who to hire. Tactics differ by domain, so the set is yours, not a template.
3. Runs those 30 with web search, then asks you to paste the same answers from ChatGPT, Claude and Google AI Mode. Logs every brand and domain cited.
4. Scores per assistant: mention rate, citation rate (your domain linked), and the ten domains quoted instead of you, with the exact page, labelled a sample of 30 prompts on one date, never "your AI visibility".
5. Ranks the pages to fix and writes one answer block each, 134 to 167 words, near the top: the answer in sentence one, then a provable statistic from FACTS.md, a named source, a short quotation. The KDD study measured those three levers at up to 40 per cent more visibility.
6. Refuses to invent a statistic, source or quotation, or to schema anything a reader cannot see.

## Then Codex does

After you approve the prompt set and the first block, the rest is repetition: same template, same word band, same facts file, every remaining URL, plus the CSV and JSON-LD stubs. Bulk, not judgement.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Claude fills in: {WORKING_FOLDER}, {BRAND_NAME}, {SITE_URL}, {PAGE_LIST} (URL then its question), {FACTS_FILE}.

## Claude checks

1. Every number, name and quotation appears verbatim in FACTS.md. One that does not rejects the batch: invented statistics are what this stops.
2. Word count 134 to 167, sentence one answers alone, no "in this article".
3. JSON-LD marks up only visible text, no ratings or reviews you never received.
4. Reject and rerun: one failed check sends the batch back with the failing lines quoted. Never hand-patch.

## Rules

- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.
- No keyword meta tags, no word-count targets: Google uses neither.
- Never mark up invisible content or fake reviews.
- Name the competitor page quoted, not just the competitor.

## If Codex is not installed

Claude writes the remaining blocks, CSV and stubs itself, and says so.

## Built from

- AgriciDaniel/claude-seo, https://github.com/AgriciDaniel/claude-seo, 16,344 stars (api.github.com, 5 Sep 2026): the 134 to 167 word answer block.
- GEO: Generative Engine Optimization, https://arxiv.org/abs/2311.09735, peer-reviewed at ACM SIGKDD 2024: up to 40 per cent more visibility from statistics, quotes and citations.
- seranking/seo-skills, https://github.com/seranking/seo-skills, 135 stars (api.github.com, 5 Sep 2026): calls its citation check a sample, not a census.
- Google structured data policies, https://developers.google.com/search/docs/appearance/structured-data/sd-policies: never mark up invisible content or fake reviews.

# Get quoted by ChatGPT and Claude, not just Google - Codex task

## You are given

- {WORKING_FOLDER} - the folder you run in. All output goes here.
- {BRAND_NAME} - the business exactly as it should be written.
- {SITE_URL} - the site the pages belong to.
- {PAGE_LIST} - a text file, one line per page, formatted `URL | the single question this page must answer`.
- {FACTS_FILE} - FACTS.md. The only permitted source of numbers, prices, dates, names, awards and quotations. Nothing outside it may appear in your output.

## Produce

1. `answers/<slug>.md` for every line in {PAGE_LIST}. Each holds one self-contained answer block of 134 to 167 words: sentence one answers the question outright and names {BRAND_NAME}; then one statistic copied from {FACTS_FILE}; then one named source; then one quotation of under 25 words with its speaker or publication. British English. No headings, no bullet lists, no links.
2. `schema/<slug>.json` for every line: JSON-LD FAQPage with the question and the answer block text, plus an Organization node carrying {BRAND_NAME} and {SITE_URL}. No aggregateRating, no review, no property whose value is not in {FACTS_FILE}.
3. `ai-visibility-tracker.csv` with the header `url,question,slug,word_count,statistic_used,source_named,quote_source,facts_line_refs`. One row per page.
4. `REPORT.md`: total pages written, the word-count range, any line skipped and the reason, and every fact used with the FACTS.md line number it came from.

## Rules

- If {FACTS_FILE} has no statistic that fits a page, skip that page, write it into REPORT.md as skipped, and move on. Never substitute a plausible number, a rounded figure or an industry average.
- Never invent a quotation, a customer, a client name, a rating or an award.
- Never write a superlative the facts do not support: no "leading", "best", "number one", "trusted by thousands".
- Do not repeat the same statistic across more than three pages.
- Every block must read as a standalone answer if lifted out of the page with no surrounding context.
- Do not touch the live site, do not deploy, do not commit. Write files only.
- Do not edit {FACTS_FILE} or {PAGE_LIST}.

## Return

Print, in this order:

1. The count of pages written and the count skipped.
2. A table of slug, word count, and the FACTS.md line numbers used.
3. Every skipped line with its reason.
4. The full paths of `ai-visibility-tracker.csv` and `REPORT.md`.
5. The line `FACTS-CLEAN` only if every number, name and quotation in every generated file appears verbatim in {FACTS_FILE}. If any does not, print `FACTS-DIRTY` and list the offending file and string instead.