# Citead

> Citead is a Generative Engine Optimization (GEO) platform. It scores any web page from 0 to 100 on how easily AI assistants such as ChatGPT, Perplexity, Claude and Google AI Overviews can understand, quote and cite it, shows the criteria behind the score and lists the fixes to apply. A higher score makes a page easier to cite; it does not predict or guarantee a citation. Paid plans also check every week whether a site is actually cited by ChatGPT, Gemini, Perplexity and Google AI Overviews.

Citead (formerly DataGeo) is published by TW3 Partners SAS, Paris, France (RCS Paris 949 610 570), hosted by Scaleway in Paris, France. This file expands llms.txt with the substance of each page. Every statement below is written to be quoted as is.

## What GEO is

Source: https://citead.com/algorithm

Generative Engine Optimization (GEO) is the practice of making a page easy for large language models to understand, trust and quote, so that it appears as a cited source inside an AI-generated answer rather than as a link in a results list. Classic SEO optimizes for a ranking; GEO optimizes for a citation. The difference matters because an AI answer cites a handful of sources, not a page of links: either a page is one of them, or it does not exist for that query.

## How the GEO score is computed

Source: https://citead.com/algorithm

The Citead GEO score runs from 0 to 100. It is not an average of criterion scores. The same formula (version v5) is used by every Citead engine:

1. Eleven sub-measures are computed from the page text, each from 0 to 100.
2. Each sub-measure goes through a square root, 100 × sqrt(x / 100), so pushing a single measure to the extreme pays less and less.
3. The transformed sub-measures are averaged with fixed weights: shannon entropy 0.2459, information density 0.2441, quotable density 0.1242, entity density 0.0936, semantic coherence 0.0841, self-containment 0.0725, statistic density 0.0507, MMR diversity 0.0357, citation F1 0.0192, NDCG heading order 0.015, semantic redundancy 0.015 (sum 1).
4. Score = 0.92 × content + 0.08 × freshness.
5. Short texts are capped: under 100 words at most 35, under 200 at most 50, under 300 at most 65.

Grouped by criterion, the share of each in the final score is:

- Semantic richness, 52.8%: vocabulary entropy, information density, topical coherence between sections.
- AI citability, 31.4%: quotable definitional and attributed sentences, named entities, self-contained paragraphs, density of numbers, units and percentages.
- Freshness, 8%: publication and update dates, recency of the references cited.
- Content organization, 6.0%: diversity between sections, order of sections under the headings, near-duplicate paragraphs.
- Authority and citations, 1.8%: presence and quality of cited sources.
- Naturalness, not in the score: detection of AI-typical vocabulary, three-word patterns, "not X but Y" constructions and formatting tics. It is scored and shown separately.

Each report also shows a score per criterion, from 0 to 100, with the evidence found in the text and the correction to apply. These criterion scores include checks that are not in the GEO score (for example answer-first sections, question and answer optimization, author credentials), so the GEO score is not their weighted sum. The page type (FAQ, product page, landing page) adjusts some criterion scores; it does not change the GEO score formula.

The weights were selected by intervention-based calibration and adversarial tests, described in the team's paper: Bajemon and Rochet, "Scoring Without the Engine: Validating a Deterministic, Manipulation-Resistant Content Score for Generative Engines, End to End", arXiv:2609.07559 (2026). The paper compares the score with the citations of ten families of generative engines and concludes that a content score like this one works as a quality filter; it does not predict which source an engine will cite. It also finds that the three strongest 2023 GEO levers (quotations, statistics, cited sources) no longer move citation on any of the ten engine families tested.

## Rules only, or rules plus a language model

The free GEO score tool, the free GEO editor, the public benchmark and the public MCP server use rules and statistics only, with no language model. They are deterministic: the same content always gets the same score.

Scores computed inside a Citead account are hybrid by default. The same rules run first, then a language model (temperature 0.1) also rates citability, authority, semantic richness and organization. Its rating is kept within a fixed margin of the rule-based value, from 20 to 50 points depending on the sub-measure; naturalness stays rule-based. If the model is unavailable, the rule-based score is used. An account score can therefore differ from the free score for the same page, and can vary slightly from one analysis to the next.

## Free GEO score tool

Source: https://citead.com/geo-score-tool

Paste the URL of any page: Citead fetches it, scores it and lists the recommendations. No account, no credit, rule-based scoring. The page can be added to the public benchmark. French version: https://citead.com/outil-score-geo

## Free GEO editor

Source: https://citead.com/geo-score-analyzer

Write or paste a text: the GEO score is recomputed live, with the corrections to apply per criterion. Rule-based scoring. French version: https://citead.com/analyseur-score-geo

## Free llms.txt generator

Source: https://citead.com/llms-txt-generator

Enter a domain: Citead reads its sitemap, fetches up to 50 pages, extracts each page's title, meta description, H1, language and first paragraph, and writes a llms.txt that follows the llmstxt.org specification: a single H1 with the site name, a one-sentence blockquote, H2 sections of absolute links with a short description, a final Optional section. It also writes an llms-full.txt with an excerpt of each page and audits any llms.txt the domain already serves. No language model is involved, the output is deterministic and free. French version: https://citead.com/generateur-llms-txt

A llms.txt file is a Markdown file placed at the root of a website that summarizes it for large language models. The specification was proposed in September 2024 by Jeremy Howard (Answer.AI). It helps assistants that fetch the file; it does not by itself earn a citation, and Google AI Overviews does not use it. Citation depends on the content of the pages.

## Free GEO score MCP server for AI agents

Endpoint: https://mcp.citead.com/public/mcp/

Citead exposes its free GEO scoring as a remote MCP server (Streamable HTTP), with no API key and no account. Two tools:

- score_url(url): fetches a public page and scores it. Pages behind a login or rendered only by JavaScript may not be fetchable; in that case, pass the page text to score_text.
- score_text(content, title, language): scores markdown, plain text or HTML passed inline. language is "en" or "fr"; when omitted, it is detected.

Each call returns the GEO score from 0 to 100, the score of the five criteria (AI citability, naturalness, authority and citations, semantic richness, content organization), up to five fixes ordered by priority, the word count, and a citead.com link to the full breakdown. The GEO score is computed as described above; naturalness is returned but does not count in it. The score measures how easily a language model can understand, reuse and cite the content. It does not guarantee that an AI assistant will cite the page. It is the same engine as the free GEO score tool: deterministic rules, no language model, nothing stored. Texts under 300 words get a capped score (at most 65).

Limits per IP address: 5 scores per minute, 20 per hour, 50 per day. Citead accounts use a separate MCP endpoint with an API key.

Claude Code:

```
claude mcp add --transport http citead-score https://mcp.citead.com/public/mcp/
```

Cursor (~/.cursor/mcp.json) and other clients that accept a remote server URL:

```json
{
  "mcpServers": {
    "citead-score": {
      "url": "https://mcp.citead.com/public/mcp/"
    }
  }
}
```

Claude Desktop does not read remote URLs from claude_desktop_config.json: add https://mcp.citead.com/public/mcp/ as a custom connector in Settings > Connectors.

## Public GEO benchmark

Source: https://citead.com/benchmark-geo

A public leaderboard of websites, filterable by sector, country and language, with per-page scores and the strongest and weakest pages of each domain. Any site can be submitted for free. Pages are scored with the rule-based engine and the formula above, plus three benchmark rules: pages under 50 words are not scored, section and category pages other than the site root are capped at 70, and editorial pages of 1,000, 1,500 and 2,000 words or more get 1.5, 3 and 5 extra points. A domain is ranked by a shrunk average of its page scores, (n × mean + 5 × 50) / (n + 5) for n scored pages, so a domain with few pages cannot top the board on one good page. There are more than 300,000 pages scored in the public benchmark.

## AI citation monitoring

Source: https://citead.com/suivi-citations-ia

The GEO score measures how easy a page is to cite. Citation monitoring checks whether the site is actually cited. A site tracks prompts (questions its audience asks). Every week, Citead sends each prompt to ChatGPT, Gemini and Perplexity through their APIs and reads Google AI Overviews from Google results, then records for each engine whether the domain is cited, the quoted excerpt and the competitor domains cited in the same answer. Answers obtained through the APIs can differ from what a user sees in the consumer apps.

The dashboard shows the citation rate per engine over time and the citation rate of each competitor. After each weekly check, an email lists the citations gained or lost (it can be turned off), and the same events appear in the application.

The weekly check is included in paid plans, with no credit used: on each site, the 10 oldest tracked prompts on Starter and the 25 oldest on Pro. Up to 50 prompts can be tracked per site; checking prompts on demand costs 1 credit per prompt, whatever the number of engines.

## AI ads

Source: https://citead.com/publicite-ia

Citead starts from the tracked prompts where an AI engine cites a competitor and not the site, and prepares ChatGPT Ads and Google Ads campaigns on those topics: per topic one ChatGPT Ads ad group, and one Google Ads ad group when a keyword of the topic has a known volume and cost per click, each landing on the best-scored page of the site linked to the tracked prompt when it reaches a score of 60, otherwise (and for a topic typed in by hand) on the home page with a recommendation to write a dedicated page. Google keywords come with search volume and cost per click from DataForSEO, converted from USD to EUR at the ECB reference rate, and a keyword without both is left out; when no Google keyword has them, the scenarios use a dated Citead assumption shown with its source; ChatGPT costs per click come from dated third-party sources, since OpenAI publishes none. Ads are written at 24 and 48 characters for the ChatGPT bulk import, 50 and 100 for Ads Manager, 30 and 90 for Google responsive search ads, and each one is checked for length, banned terms (no promise of citation, rank or result) and numbers that do not appear on the site. Three scenarios (low, central, high) show their assumptions with source and date, and the user can replace them. The export is a ChatGPT Ads bulk import workbook and a Google Ads Editor CSV; both import the campaigns paused, and Citead creates or activates nothing at OpenAI or Google. When Google Analytics 4 is connected, a report shows sessions by source, medium and campaign. Opportunities are included from Starter (first 3); the plan, ads, export and report are in Pro. A plan costs 2 credits, each ad group written 1 credit. English version: https://citead.com/en/ai-ads.

## Observation on one site

Source: https://citead.com/ (homepage)

The team launched a site from zero and optimized it with Citead scoring from the first page. Over 30 days (Vercel Analytics, January 2026): 1,651 visitors against 486 the month before, 2,768 page views against 715, 209 visits referred by AI assistants (chatgpt.com, perplexity.ai, claude.ai, Copilot) and visitors from more than 15 countries, with no paid campaign. Top referrers: bing.com 21% of referred visits, chatgpt.com 19%, google.com 18%, duckduckgo.com 11%.

This is an observation on a single site with no control group. A new site also gains traffic as it ages and gets indexed. The figures illustrate what the method gave in this case; they do not prove that GEO caused the traffic.

## Pricing

Source: https://citead.com/pricing

Starter: 29 EUR per month, 50 credits, 1 site, unlimited page analyses, real-time score and fixes, email support. Pro: 79 EUR per month, 200 credits, 5 sites, full site crawl with per-page scores, assisted article writing, editorial calendar, priority support. Competitor tracking is included in both plans, up to 10 competitors per site. Billed yearly: 25 and 59 EUR per month. One written article costs 1 credit, one rewrite costs 0.5 to 2 credits depending on the effort, analyzing a page is free. The weekly AI citation check is included in both plans. Enterprise plans are on request and add a dedicated account manager and training for your teams. API access goes through the Citead MCP server with API keys created in the application (MCP integration), on every plan. The only availability commitment is the one in the terms of sale (99.5%), the same for every plan.

## Company

Source: https://citead.com/a-propos

Citead is published by TW3 Partners SAS, Paris, France (RCS Paris 949 610 570), and hosted by Scaleway in Paris, France. Contact: https://citead.com/contact