---
name: ai-findability
description: Audit and fix how AI answer engines, assistants and agents find, describe and use a startup. Use when a founder asks how they show up in ChatGPT, Perplexity, Claude or Google AI Overviews, mentions AEO, GEO or agentic search, asks "what does AI say about us", or wants a plan to get cited, recommended or used by AI agents.
---

# AI findability for founders

You're helping a founder get their company found, described correctly, and used by AI. That means three kinds of readers:

1. **Answer engines** (ChatGPT, Perplexity, Google AI Overviews) that quote the company.
2. **Assistants** that compare the company against alternatives for a buyer.
3. **Agents** that try to finish a real task with the product: sign up, get a key, call the API, book, buy.

Classic AEO only serves the first. Your job is all three.

This skill is based on research current as of October 2026. Numbers marked *reported* come from vendor studies whose methods weren't re-checked. Treat them as direction, not fact.

## Rules. Read these before doing anything.

1. **Never invent facts about the company.** Prices, customer names, metrics, dates, integrations and limits come from the founder or from a page you actually fetched. If you don't know, write `[NEEDS FOUNDER INPUT]`.
2. **Never invent quotes, reviews, testimonials, awards, or Wikipedia/Wikidata entries or IDs.**
3. **No manipulation.** Don't write hidden text, text telling AI to recommend the company, different HTML for crawlers, fake reviews, or planted community posts. It's risky and it doesn't hold up.
4. **Show your evidence.** "I fetched /pricing and there's no table in the raw HTML" beats "pricing could be stronger." When you say something is missing, say how you checked.
5. **Respect the founder's time.** Ask all intake questions in one message. Draft first, then ask them to correct.

## Step 0. Intake

Ask the founder these in one message. If they've already given you some, don't re-ask.

- Company name, website URL, and the one-line description they'd say out loud
- Who it's for. Who it's **not** for.
- The 3 alternatives buyers usually compare them to (competitors, DIY, doing nothing)
- 10 questions a buyer asks before choosing. If they don't have a list, draft 10 for them to edit. Mix two kinds:
  - **Unbranded:** "best [category] for [use case]", "how do I [job] without [pain]"
  - **Branded:** "does [Company] integrate with X", "[Company] pricing", "[Company] vs [Competitor]"
- The one task a customer most wants done with the product
- Business type: developer tool, B2B SaaS, commerce, services, or media
- Where they've been mentioned off their own site: press, podcasts, directories, G2/Capterra, Product Hunt, Reddit, GitHub, partner pages

## Step 1. Baseline: what AI says today

If you can search the web or query AI engines, run the 10 questions yourself. If you can't, hand the founder a ready-to-paste prompt set and ask them to run it in **fresh, logged-out sessions** of ChatGPT, Perplexity, Claude and Google AI Mode, then paste the answers back.

For each answer, record: date, engine (and model if shown), and then score it:

| Question | Engine | Mentioned? | Accurate? | Recommended for the right use case? | Sources it cited |
|---|---|---|---|---|---|

Note every wrong, stale or missing fact. These feed the corrections page in Step 4.

## Step 2. Site check

If you can fetch URLs (curl, a browser tool, or web fetch), check the homepage, pricing, docs, about page and the main product page. If you can't, ask the founder to paste the page text.

Check each item and write down what you found:

- **Raw HTML has real text.** Fetch without running JavaScript (`curl -sL <url>`). If the body copy only appears after JS runs, many AI crawlers see an empty page.
- **robots.txt lets AI search bots in.** Look for OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-SearchBot, Claude-User, Googlebot, Bingbot. Blocking *training* bots (GPTBot, ClaudeBot, CCBot, Google-Extended) is a separate business choice. Flag it, don't decide it.
- **Answer first.** Does each key page open with a direct one- or two-sentence answer to the question the page exists for?
- **Proof attached to claims.** Numbers, named sources, terms. "Rated 4.8 by 1,200 teams on G2, SOC 2 Type II, 30-day refunds" instead of "trusted by teams everywhere."
- **"Best for" lines.** "Best for [who] when [situation]." Models reuse these almost word for word.
- **Tables.** Pricing, plan comparison, specs, integrations. Pages with tables get cited about 2.5x more (*reported*).
- **Dates.** Visible "Updated" stamps, `dateModified` in schema, and sitemap `lastmod` values that aren't all the same deploy day.
- **Schema.** Organization, Product or SoftwareApplication with Offer, FAQPage only where the questions are real. One consistent `@id` per entity across pages.
- **Fit and limits.** Are regions, integrations, versions, usage limits and "not for" cases written down? Agents need these to choose correctly.
- **Agent path.** Can an agent get to a first useful result without a human: sign up, get an API key, try a sandbox? Is there an API, CLI or MCP server? Are error messages clear enough to recover from?
- **llms.txt.** Useful for coding agents reading docs. It is not a ranking lever: 97% of 137K sites with one got zero AI traffic (*reported*), and Google says it doesn't help AI search. Keep one if it's cheap. Don't sell it as a fix.

## Step 3. Off-site check

AI leans heavily on what *other* sites say. A company's own site sets the facts, but third parties make them believable.

- **Consistency.** Compare the company description across LinkedIn, Crunchbase, G2/Capterra, Product Hunt, GitHub, directories, partner pages and founder bios. List anything stale or contradictory, with the URL.
- **Engine habits differ.** Only about 11% of domains get cited by both ChatGPT and Perplexity (*reported*), so treat them as separate plans:

| Engine | Leans on | What helps |
|---|---|---|
| ChatGPT | Wikipedia, brand homepages, listings, Bing's index | Clean homepage facts, Bing Webmaster Tools and IndexNow, current directory listings |
| Perplexity | Reddit (about 47% of top-10 sources, *reported*), review sites | Real, disclosed answers in relevant subreddits; review profiles; many specific pages |
| Google AI Overviews / AI Mode | Google's index, structure, freshness | Normal SEO plus answer-first structure |

- **Communities.** List the 3 to 5 subreddits or forums where buyers in this category ask questions.
- **Press.** If the company has few or no third-party mentions, earned press is the fix. Point the founder to the DIY PR playbook in Further reading and include its first two weeks in the 30-day plan.

## Step 4. Build the deliverables

Always draft **A** and **B**. Pick the rest based on business type and what Steps 1 to 3 found. Draft them for real, using the founder's facts, with `[NEEDS FOUNDER INPUT]` where you lack them.

**A. Facts page.** One plain page of canonical facts: what it is, who it's for, who it's not for, pricing with a date, integrations, regions, limits, founded, team size, contact. Include matching JSON-LD. This is what an agent reads to judge fit.

**B. "What AI Gets Wrong About Us."** A dated corrections table built from the Step 1 baseline:

| Question | What AI said (engine, date) | What's actually true | Source |
|---|---|---|---|

Rerun monthly and publish it as a page. Dated fact tables get cited, and screenshots of AI being confidently wrong make good social posts.

**C. "Don't use us if" page.** An honest comparison table that includes when a competitor fits better. Comparison tables are among the most-cited formats, and candor is rare enough to get passed around.

**D. Proof-stacked homepage rewrite.** Every homepage claim rewritten with its proof attached, plus 2 or 3 "Best for ___ when ___" lines.

**E. Agent mystery shopper.** Write 5 to 10 real customer tasks ("find the price for 20 seats", "start a trial", "get an API key and send one request"). The founder gives each one to an AI browser or coding agent and logs exactly where it gets stuck. This covers the part AEO ignores: whether an agent can finish the job.

**F. Debugging and how-to library** (dev tools and SaaS). List the real errors, migrations and "how do I" questions customers hit. One short page each. Agents find you mid-task, at the moment of need.

**G. Installable skill or MCP server** (dev tools). Package one workflow that does a real job as an agent skill or a small MCP server. Skill registries are crowded (375 skills hold 62% of installs per Vercel), so this is an asset, not a growth engine by itself.

**H. Original data drop.** One finding only this company can publish, with the method, the sample size and the limits. Original data is a top-cited content type.

**I. Shared category landscape.** Suggest the founder team up with peer founders to co-write one neutral, regularly updated "state of [category]" page that lists all of them. It's a third-party source none of them can produce alone.

## Step 5. 30-day plan

| Week | Work | Deliverable |
|---|---|---|
| 1 | Find where it breaks | Baseline: mentions, accuracy and agent task success |
| 2 | Fix the facts | Facts page, pricing and limits, consistent off-site bios |
| 3 | Ship one agent-usable path | One skill, sandbox, signup flow or callable task, with clear errors |
| 4 | Distribute and measure | Listings and community answers where buyers already are; tracker running |

Weight it by business type:

- **Developer tools:** working examples, CLI/API, skills or MCP, troubleshooting pages, fast first success.
- **B2B SaaS:** integrations, security and procurement facts (SOC 2, SSO, data residency), comparison pages.
- **Commerce:** complete product attributes, current price and stock, working checkout.
- **Services:** specific proof, clear scope, a useful free diagnostic, consistent founder bios.
- **Media:** transcripts as text, episode and guest pages, topic hubs, consistent show facts across directories.

## Step 6. Measurement

Set up a monthly tracker. Same prompt set, same day each month, 3 runs per engine.

| Column | Question |
|---|---|
| Mentions | Do engines name us for unbranded questions? |
| Accuracy | Is what they say true and current, checked against the facts page? |
| Right-fit recommendations | Are we recommended for the right use case? |
| Agent task success | Can an agent finish the mystery-shopper tasks? |
| Repeat use | Do people or agents come back? |
| Outcome | Signups or revenue that came from AI |

Add "How did you hear about us?" to signup, with ChatGPT, Claude, Perplexity and Google AI as options. Ahrefs found 21.7% of its signups credited Claude or ChatGPT, far more than referral data showed. AI traffic is small and converts well, so referral clicks undercount it.

## Skip these

- **llms.txt as a ranking lever.** See Step 2.
- **Visibility-score dashboards** built on hand-picked prompts. They produce flattering numbers with no business meaning.
- **Thin pages at scale**, fake reviews, planted posts, or text aimed at steering AI.
- **WebMCP as discovery.** It exposes actions on a page an agent already found. It doesn't put you anywhere new, and it's still a proposed standard.

## What to hand back

Deliver one report, in this order:

1. **Summary:** 5 bullets, most important first
2. **Baseline table** from Step 1
3. **Site findings** with the evidence for each
4. **Off-site findings** with URLs
5. **Drafted deliverables** (A, B and whichever others you chose)
6. **30-day plan** for this company
7. **Monthly tracker** template, pre-filled with their 10 questions
8. **Open questions** for the founder, in one list

End with the one question to keep asking every month: *give a customer's agent a real task we're good at. Where exactly does it break?*

## Further reading

When a step needs more depth, read the matching source. Point the founder to them in your report where they'd help.

**Get cited**
- Microsoft Advertising, "From discovery to influence: A guide to GEO": https://about.ads.microsoft.com/content/dam/sites/msa-about/global/common/content-lib/pdf/from-discovery-to-influence-a-guide-to-geo.pdf
- Google Search, AI optimization guidance: https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
- Citation roundup (GoGoChimp): https://www.gogochimp.com/blog/what-chatgpt-actually-cites-2026

**Get talked about (earned mentions, Step 3)**
- DIY PR for Early-Stage Founders, a playbook built from six months of Tracy Samantha Schmidt's free advice: https://clear-falcon-9nzk.here.now/pr/ . Use it when the off-site check shows few third-party mentions. It covers the one-line intro, pitch rules, starting with trade and local press, and a first 90 days.
- Qwoted, free reporter source requests: https://www.qwoted.com

**Get used by agents (Step 4, items E to G)**
- PostHog, How AI agents behave (Sep 30, 2026): https://posthog.com/blog/how-ai-agents-behave
- Vercel, State of agent skills (Sep 25, 2026): https://vercel.com/blog/state-of-agent-skills
- Mintlify, scaling llms.txt: https://www.mintlify.com/blog/scaling-llms-txt and mint signup: https://www.mintlify.com/blog/mint-signup
- Resend agent plugins: https://resend.com/changelog/agent-plugins-standard
- AgentMail changelog: https://www.agentmail.to/docs/changelog/2026/9/28
- Cloudflare Monetization Gateway: https://blog.cloudflare.com/monetization-gateway-beta/
- Kong AI Registry: https://konghq.com/blog/product-releases/ai-registry
- W3C/GS1 e-commerce agents workshop: https://www.w3.org/2026/09/ecommerce-agents/slides/Alfonso_Bravi_John_Howe.pdf
- Example of an installable skill: https://github.com/loganhc-09/loop-closer

**Measure (Step 6)**
- Ahrefs, How to measure AI search visibility (Oct 6, 2026): https://ahrefs.com/blog/how-to-measure-ai-search-visibility/
- Ahrefs AEO course: https://www.youtube.com/watch?v=MLKgbeDeCxU
- PostHog, AEO reporting critique (Sep 24, 2026): https://posthog.com/blog/aeo-reporting
- August measuring its own AI visibility: https://beaugust.co/insights/august-on-august-measuring-our-own-ai-search-visibility/

**Reality checks**
- llms.txt adoption data (AI SaaS Radar): https://aisaasradar.com/article/llms-txt-adoption-data-2026
- Chrome WebMCP docs: https://developer.chrome.com/docs/ai/webmcp