Overview
Every report answers three questions: Do AI assistants know you? Do they recommend you? Can they read your site? Each gets a 0–100 score, combined into one AEO score:
The free audit asks one buyer question to small, no-web-search versions of ChatGPT, Claude and Gemini (GPT-6 Luna, Haiku 5.5, Gemini 3.5 Flash-Lite) and the three open models, and takes under a minute. Weekly scans ask 25 (Starter), 100 (Growth), 150 (Managed) questions to every assistant — up to 906 live AI calls per scan — and run in the background.
The assistants we ask
Two kinds of engines, because they answer different questions:
| Assistant | Model | How it answers |
|---|---|---|
| ChatGPT | ChatGPT GPT-6.1 Sol (OpenAI) | Live web search |
| Claude | Claude Sonnet 5.5 (Anthropic) | Live web search |
| Gemini | Gemini 4 Argon (Google) | Training data only |
| DeepSeek | DeepSeek V4 Pro (DeepSeek) | Training data only |
| GLM | GLM 5.3 (Zhipu AI) | Training data only |
| Kimi | Kimi K2.6 (Moonshot AI) | Training data only |
ChatGPT and Claude are called through their makers’ official APIs with web search turned on — the same way their consumer apps look things up before answering. They show what an assistant finds and cites about you today. Calls are routed through Cloudflare AI Gateway. They run on every weekly scan and refresh at most every 3 days on Starter, 2 days on Growth and daily on Managed; on-demand rescans in between use the other engines.
Gemini is called through Cloudflare AI Gateway as well, but without web search (Google’s search grounding is not available through the gateway), so it answers from training data.
DeepSeek, GLM and Kimi are open-weight models run on Cloudflare Workers AI without web access. They show what models have learned about you from the public web — the slow-moving layer that your content, PR and documentation shape over months.
We call the APIs, not the consumer chat apps. App answers can differ slightly because of personalisation, location, memory and app-specific instructions; we never present API answers as screenshots of the apps.
How scores work
Recognition (35%)
Each assistant is asked what your company is. A separate analyst model (Llama 3.3 70B) fact-checks every answer against the description and claims on your own homepage and returns whether the company was recognized and an accuracy rating from 0–100. Per assistant: 0 if not recognized, otherwise 40 + 0.6 × accuracy. The score is the average across assistants.
Recommendation (35%)
Each assistant answers every buyer question. We detect whether your brand or domain is named and at which list position. Per answer: 1.0 if you are in the top 3, 0.75 if mentioned lower, 0 if absent. The score is the average across all answers × 100. We also count every other brand named, which becomes the “who AI recommends instead” list and your competitor benchmark, and we break results down per assistant.
Readiness (30%)
A deterministic audit of your site, worth 100 points in total — see the table below. No AI is involved in this score.
What we ask
Prompts are fixed so results are comparable week to week:
Recognition
What is {Brand} ({domain})? In 3-4 sentences describe what the company does, who it is for and how it is positioned. If you are not familiar with this specific company, reply "I'm not familiar with {Brand}." and do not guess.
Recommendation
{Buyer question} Recommend the 5-8 best options as a numbered list, one line each: name — why.
On your first scan we generate your plan’s number of buyer questions from your homepage, spread across five angles: best-of lists, problems and use cases, alternatives and comparisons, company size or industry, and buying criteria such as pricing, integrations and compliance. Questions never mention your brand. They are saved to your dashboard so every weekly scan asks the same set; you can edit, replace or extend them any time.
Sources AI cites
When ChatGPT and Claude search the web, they return the pages they used. We count which sites they cite across all answers and how often your own site is among them. Review sites, listicles and comparison pages that appear here are where a fair mention of your company changes AI answers fastest.
Readiness checks
| Check | Points |
|---|---|
| Key AI crawlers allowed in robots.txt (GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended) | 25 |
| Server-rendered text content (600+ words for full points) | 15 |
| Structured data: Organization / Product / SoftwareApplication (10) and FAQPage (5) | 15 |
| /llms.txt published | 10 |
| Meta description of 50–180 characters | 10 |
| Descriptive title, single H1, XML sitemap, HTTPS + canonical, Open Graph | 5 each |
We also report the robots.txt status of 9 AI crawlers: GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot, Google-Extended, Applebot-Extended, meta-externalagent, CCBot.
Limitations
- AI answers are probabilistic. The same question can produce different lists. Many questions, six assistants and fixed prompts reduce the noise; the weekly trend is the signal.
- API answers are close to, not identical with, the apps. Consumer apps add personalisation, memory and location. Google AI Overviews and AI Mode have no public API, so we do not measure them.
- Open models have training cut-offs. Changes you ship today show up in their knowledge over months; web-search assistants pick them up much faster.
- Brand matching is literal. We match your brand name and domain. Very generic names can occasionally be over- or under-counted; full answers are always shown so you can verify.
- We read only your homepage plus robots.txt, llms.txt and sitemap. Pages behind logins or bot protection are invisible to us — and usually to AI crawlers as well.
Your data
We only fetch public pages. The questions we send to AI providers contain your brand name, domain and category — never personal data. Reports are stored in Cloudflare KV and visible only through your private dashboard. Free audits are cached for 24 hours and then deleted. Details in our Privacy Policy.