Citewake

The index of what AI recommends.

Each month, Citewake asks ChatGPT, Claude, Gemini and Perplexity the questions buyers ask. It keeps every answer, and measures who they recommend, who they cite, and which sources list whom.

The first index publishes early 2027. The method is public now.

Illustration of the method, not real results. One buyer's question goes to four AIs, each searching the web on its own. One reader goes through all four answers side by side and marks every brand by its role.

AI names three to five brands. Nobody can see why, or track it.

Ask an assistant for the best tool for a job and it gives you a short list, mostly the biggest names. Whether a brand is on that list now shapes who gets considered.

But each answer is private, the sources behind it shift, and tomorrow's answer may differ. There is no public record of what the AIs recommended, which pages they relied on, or how that changes month to month.

Citewake is that record. Fixed questions, asked the same way on a schedule, with every answer kept as evidence.

How it works

Every observation passes through three separate compartments, each with its own rules.

  1. The prompt

    The question goes to each AI exactly as written, through its official API. No memory, no custom instructions, no system prompt. Answers are never cached or reused; every observation is a fresh request.

  2. Each AI's own search

    Each AI searches the web with its own search, reads what it finds and answers. The search is never shared and never substituted, because which pages each AI finds and trusts is a large part of what Citewake measures.

  3. The analysis

    One judge, the grand master, reads all four answers to a question side by side, with their citations. It resolves each brand to the same entity in all four and holds every AI to one standard.

The grand master classifies. It doesn't judge.

For every brand in every answer it records the exact words, and the role: recommended, a neutral mention, or negative. A brand is kept only if its name, alias or domain appears in the answer text and the quoted words are found verbatim. It never adds a brand that isn't there and never adds an opinion.

Reader
Claude Opus 5.5, for all four AIs' answers
Cited pages
Read by Claude Haiku 5.5 under the same rulebook; uncertain pages and a random 10% go to the grand master
Independence check
Accuracy reported separately for each AI's answers, Claude's included

Three measures, each with a fixed denominator

Each one counts one thing, so a young brand has real numbers that move before it is ever recommended. Every rate is shown per AI, per tier and overall, with a 90% interval.

Recommendation Rate

How often the AIs positively recommend the brand.

Valid answers that positively recommend the brandValid answers in the benchmark

Own-site Citation Rate

How often answers cite the brand's own website.

Valid answers that cite a URL on the brand's own domainsValid answers in the benchmark

Cited-page Presence

How often the third-party pages the AIs cite list the brand.

Cited third-party pages, in their fetched copy, that list the brand positivelyUsable cited third-party pages in the benchmark

Supporting, and labelled as such: Neutral and Negative Mention Rates, kept apart from recommendations; Competitive Share, which is relative and can fall when rivals gain; and Reported Retrieval Presence, shown per AI only where that AI's API reports it.

No combined score yet. A single 0–100 number is held until the pilot shows that combining the measures adds information beyond them, and that its weights hold on data they weren't tuned on.

Categories

Each category is defined by the job a buyer wants done, with stated boundaries. The first three are markets we know well, so every classification can be checked by hand. More follow as the index grows, each indexed when there's demand for it, and every lookup on this page counts.

  • In preparation

    AI search visibility tools

    Also called GEO or AEO tools

    The jobSee and improve how AI assistants mention, recommend and cite a brand.

    Doesn't belong

    • SEO suites where AI visibility is a feature, not a product
    • AI writing tools that claim to optimise for AI without measuring answers
    • Agencies and consultancies
    • One-off audits with no ongoing measurement
  • In preparation

    Link-in-bio tools

    The jobGive a person or business one link for their social profiles that opens a page of their links, content or products.

    Doesn't belong

    • General website builders
    • URL shorteners: one link to one destination
    • Shops where the link page is secondary
    • Features inside a social network
  • In preparation

    Live race tracking

    The jobLet spectators, organisers and supporters follow participants' live positions on a map during a race, ride, run or other moving event.

    Doesn't belong

    • Training apps where live sharing is a side feature
    • Timing companies that publish results only
    • Everyday location sharing
    • Live scores for ball or court sports
    • Fleet or asset tracking

The full public directory of categories is coming.

Scored questions
120 per category, in three tiers of 40: head, mid and tail
AIs
ChatGPT, Claude, Gemini and Perplexity, each with its own live web search
Market
Location unspecified
Edition
Monthly. Every question asked of every AI once per edition
First edition
Early 2027

Brands enter only through evidence: when an AI names them, or a cited page lists them. Nobody can submit, request or pay for inclusion. Products made by our team are in these categories, measured by the same rules with no adjustment. How we handle that.

Principles

The rules the index is held to, written into the method.

  • Every answer kept

    The full answer, its citations, and a snapshot of every cited page, forever. Any published number can be recomputed after the live pages change.

  • A public, versioned method

    The scored questions and the method are published. Any change makes a new version, dated in the changelog. Numbers are compared within a version only.

  • No paid inclusion

    Nobody can buy a place, a number or a correction. Our own brands are measured by the same code, with no manual adjustment.

  • Corrections added, not overwritten

    The original stays. A corrected figure is published beside it with the date, the cause and the evidence.

Check a website

See where a site stands in the index. Before the first index publishes, a lookup tells us which categories to index next.