Detekia GEO Score: The 6 Criteria That Matter to Get Cited by AI
Detekia GEO score methodology: 6 criteria, a readability prerequisite, their scoring rules and the September 30, 2026 study of 317 pages that calibrated them.
Contents
The Detekia GEO score is not a black box. Behind every score from 0 to 100 there are 6 specific criteria and one prerequisite, measured automatically on the page analyzed. How each criterion is evaluated, why it matters, and how to improve it: that is what this article is about.
Since September 30, 2026, the scoring grid has been based on our calibration study of the same date: 317 pages from 40 websites frequently cited in the AI visibility audits run by Beeleven, where pages actually cited by Perplexity, ChatGPT or Claude were compared with uncited pages from the same site, on the same topics. It also builds on academic work in GEO, notably Aggarwal et al., "GEO: Generative Engine Optimization" (Princeton, presented at KDD 2024).
How to interpret your overall GEO score
The official, up-to-date Detekia score methodology
Overview: 6 criteria and one prerequisite
The score starts with a prerequisite, then adds up 6 criteria grouped into three families:
- "Readable by AI" prerequisite (no points): AI search crawlers allowed, no noindex, content present without JavaScript. If it is missing, the score is capped at 40.
- Content: Answer structure (40 pts), Content depth (15 pts), Evidence & data (15 pts)
- Freshness: Freshness (15 pts)
- Identity & reputation: Entity clarity (5 pts), External signals (10 pts)
Each criterion is scored in points, up to its maximum. The overall score is the sum of the 6 criteria, out of 100, without normalization. The weights come from our calibration study: answer structure weighs the most, because it best separated cited pages from uncited ones. The calculation is deterministic: no AI is involved, and the only outside call is Wikidata's public API.
The prerequisite: readable by AI
Before any calculation, Detekia checks three conditions:
- AI search crawlers are allowed in robots.txt: OAI-SearchBot and ChatGPT-User for ChatGPT, PerplexityBot and Perplexity-User, Claude-SearchBot and Claude-User, Googlebot, Bingbot. Training crawlers (GPTBot, ClaudeBot, Google-Extended) do not count, since they are not used to cite your pages in real time (OpenAI documentation);
- the page is not set to noindex;
- its content is present in the HTML, without running JavaScript.
This prerequisite earns no points. If any condition is missing, the total score is capped at 40, however good the content: a page AI crawlers cannot read cannot be cited. In our study, accessibility was at its maximum on almost every page: it did not tell cited and uncited pages apart, hence its role as a lock rather than a points criterion.
The 6 criteria in detail
What it is: How well the page reads as a series of answers AI engines can reuse as is.
How it's measured:
- short answer (15 to 70 words) right after H2 and H3 headings, in proportion to the headings: 60% of headings or more → 14 points, 40% → 10, 20% → 6, at least one → 3;
- question-format headings: 3 or more → 9 points, one → 4;
- data table → 8 points (otherwise, at least two lists → 3);
- paragraphs of 20 to 80 words: 5 or more → 4 points, 2 or more → 2;
- hierarchy: one H1 and at least two H2s → 5 points.
Why it matters: This is what most clearly set cited pages apart in our study: short answers under headings, question headings, tables. Short answers placed under a heading are also among the traits ChatGPT quotes most (Search Engine Land).
How to improve:
- Put a two- or three-sentence answer under each H2, before the details
- Rephrase at least three headings as questions your customers ask
- Present comparisons and data in a real HTML table
- Break long blocks of text into short paragraphs, with a single H1
What it is: How much useful content the page offers on its topic.
How it's measured: Word count of the main content (main or article tag, otherwise the whole page). Thresholds depend on the page type. Homepage: 1,500 words → 15 points, 800 → 10, 400 → 5, 150 → 2. Other pages: 3,000 words → 15, 2,000 → 12, 1,200 → 9, 600 → 5, 300 → 2.
Why it matters: In our study, cited pages were on average more complete than uncited pages from the same site. A page that covers its topic fully offers more passages to reuse.
How to improve:
- Cover the questions related to the main topic (definitions, steps, examples, edge cases)
- Add a frequently asked questions section
- Put the content in a main or article tag so it stands apart from menus and footer
What it is: Figures and sources that make your claims verifiable.
How it's measured:
- figures: 10 or more → 6 points, 5 → 4, 2 → 2;
- external sources linked in the content (excluding social networks): 3 or more → 5 points, at least one → 3;
- attributions ("according to", "a study", "source:"): 2 or more → 4 points, one → 2.
Why it matters: Adding statistics and citing sources are among the methods that most increase visibility in generative answers (Aggarwal et al., KDD 2024).
How to improve:
- Replace vague wording with precise figures
- Link each statistic to its primary source, in the text
- Attribute data explicitly ("according to…", "a study by… found")
Full technical guide: robots.txt and llms.txt for AI bots
Schema.org and AI: the 5 priority schemas for GEO
What it is: How recent the content is, and whether that is visible.
How it's measured: Age of the most recent date on the page (structured data or visible text): 90 days or less → 9 points, one year → 6, two years → 2. Plus 3 points if a date is visible to readers and 3 points if the current or previous year appears in the content.
Why it matters: A recent update date was among the signals that set cited pages apart in our study. Pages left without an update for more than three months are over 3 times more likely to lose AI visibility (AirOps, 2026).
How to improve:
- Review key pages at least once a quarter, and show it with an "Updated on" mention
- Fill in datePublished and dateModified in the Article schema
- Mention the current year when the topic calls for it (figures, regulation, trends)
What it is: How easily an AI engine can tell who is speaking on the page.
How it's measured: Brand name in the title tag → 1 point, in the H1 or the first 120 words → 2 points; meta description of 70 to 320 characters → 1 point; organization named in structured data → 1 point.
Why it matters: This is a hypothesis our study did not confirm: these signals did not set cited pages apart. The criterion stays, with a low weight, because it costs almost nothing to fix.
How to improve:
- Put the brand name in the page title and in the first sentence
- Write a meta description that says what the company does
- Declare the organization in structured data (Organization schema)
What it is: What exists about your brand outside your site, checked at the source.
How it's measured: A Wikidata entity whose official website (property P856) is your domain → 5 points; a French or English Wikipedia article → 2 points; customer reviews or testimonials on the page → 3 points. The Wikidata lookup uses its public API, and the result is kept for 30 days per domain.
Why it matters: AI engines rely on third-party sources to identify a brand and judge its reputation. About 85% of brand mentions in AI answers come from third-party pages (AirOps, 2026).
How to improve:
- Create or complete your company's Wikidata entry, with your domain as the official website
- Show customer reviews or testimonials on key pages
- Earn mentions on third-party sites (media, comparison sites, industry directories)
How the criteria add up
The overall score is the sum of the points of the 6 criteria, out of 100. The prerequisite acts as a lock: if the page is not readable by AI, the score is capped at 40, even if the content would earn more.
Detekia recognizes four page types: the homepage (site root), the article (Article, BlogPosting, FAQPage… structured data, or a /blog, /guide… URL), the product or service page (Product, Offer or Service structured data) and other pages. Only the depth thresholds change with the type: a homepage does not need as many words as an article.
Recommended optimization order:
- Lift the cap (prerequisite): AI crawlers allowed, no noindex, content without JavaScript
- Structure your answers (Answer structure): the biggest pool of points
- Date and update (Freshness): quick points to earn
- Add figures and sources (Evidence & data): continuous improvement
- Complete the content (Content depth): editorial work
- Work on identity and reputation (Entity clarity, External signals): long-term work
Why the scoring grid changed on September 30, 2026
The former score had 7 criteria. Our September 30, 2026 study of 317 pages put it to the test: the cited page had the best score in 72% of same-site comparisons, with an average gap of 5.5 points, half of which was explained by page length. With the current grid, applied to the same pages, it is 80% of comparisons, with an average gap of 12.6 points.
What was removed, and why:
- Authority & E-E-A-T: identified author, Contact and About pages and Organization schema did not set cited pages apart. The brand name is still measured, with a low weight, in Entity clarity.
- Editorial neutrality: it worked the other way. "Best…" comparison pages are cited a lot, and the former score penalized them.
- AI accessibility: almost every page got the maximum. It becomes the prerequisite that caps the score.
- External presence: social media and sameAs links made no difference. It is replaced by External signals, checked on Wikidata, Wikipedia and customer reviews.
- Citability, Verifiability, Freshness: their useful signals are kept and reweighted in Answer structure, Content depth, Evidence & data and Freshness. Expert quotes and up-to-date copyright, with no measured effect, no longer count.
This study has limits: it covers large French websites in a few industries, it measures a correlation and not a cause, and pages protected against bots could not be read. The current grid was evaluated on the pages used to build it: it will be revalidated on the next AI visibility measurements. Another finding: on these topics, 79 to 89% of citations went to third-party sites (comparison sites, media, public websites) rather than brand websites.
What the score doesn't measure
The Detekia GEO score measures potential citability. It does not measure:
- Whether you are already cited: the citation test included in the Detekia score gives a first look (5 queries on an OpenAI model); to find out what ChatGPT, Gemini, Claude and Perplexity actually answer in your market, that is the purpose of the audit run by Beeleven, the agency that built Detekia
- The quality of the substance: a factually wrong but well-structured article can get a good technical score
- Topic coverage: a site with one excellent article vs. a site with 50 average ones
- Query popularity: being citable on a topic nobody searches for won't bring traffic
That's why the GEO score should be read as a citability potential, not a guarantee. A complete strategy combines technical optimization (GEO score) with editorial strategy (topics to cover) and presence on third-party sites, which receive most citations.
Analyze your GEO score for free
Frequently asked questions about the methodology
Is the GEO score valid for all types of websites?
The methodology was designed for B2B and B2C sites with editorial content (blogs, guides, product pages). It is less relevant for pure web applications (SaaS with no public content) or highly technical sites without a general audience.
How often should I re-evaluate my score?
After every significant optimization (technical overhaul, new articles, schema updates), and at least once per quarter. LLM algorithms evolve, and what is optimal today may change.
Does the GEO score replace the SEO score?
No, the two scores measure complementary things. Strong SEO (domain authority, backlinks, Google rankings) does not count directly in the GEO score, even if a well-known brand is more likely to have a Wikidata entry (External signals criterion). And pages that rank very well on Google can still have a poor GEO score if their content is not extractable by AI engines.
SEO vs GEO: key differences and how to combine them
Is your site visible to AI engines? Test it for free.
Analyze my site for free