AI Score

The AI Score grades readiness, which is a different object from performance. It answers a hypothetical: if a model needed the fact your page states, how much friction stands between the model and quoting you? Friction is crawl access, chunk quality, structured data, and whether anything external suggests you are worth trusting. All of it is measurable before a single citation exists, which is the point — it gives you something to move on day one instead of waiting a fortnight for retrieval layers to catch up.

The metric it gets confused with first is Domain Authority or Domain Rating, the third-party scores from link-graph tools. Those are link-based predictions of ranking, built almost entirely from who points at you. The AI Score is mostly not about links: authority and freshness together are one pillar of four, and the heaviest pillar by a wide margin is whether your own content survives being chunked. A site can carry a high DR and a mediocre AI Score simply by writing in long, context-dependent paragraphs.

The second confusion is with citation share, and here the distinction is leading versus lagging. Citation share is the outcome: how often you actually appear across tracked queries and platforms. The AI Score is the input you control. Moving the score does not move citations by itself, and a site can hold a strong score for weeks before the numbers follow, because each retrieval layer re-crawls on its own schedule. Read them together or you will misdiagnose both.

Forty-seven checks are distributed across four pillars with fixed weight envelopes: content extractability takes 40 points, and technical readiness, schema coverage and authority & freshness take 20 each. Every check resolves to pass, warn or fail, with partial credit allowed on the spectrum checks — chunk density is scored on a curve rather than a threshold, because a page with one extractable span per 500 words is meaningfully different from one with none.

The scan that feeds it completes in 30 to 60 seconds and runs three parallel tracks of 17, 18 and 12 checks — technical readiness, content extractability, and citation telemetry. Weights are re-fit quarterly against the tracked cohort and clipped back into the pillar envelopes, so the pillar arithmetic stays stable even when individual check importance shifts. That clipping is deliberate: a score whose meaning drifts every quarter is not a score, it is a mood.

Read the bands rather than the digits. 85 and above is the top decile — you are a default source in your category. 70 to 84 is solid, with long-tail queries still slipping. 50 to 69 is average, and it is where two or three high-lift fixes usually move twenty points at once because the failures are structural rather than fiddly. Below 50 means models effectively cannot use your site, and the top five fixes in the report are ordered by expected delta for exactly that case.

The headline number tells you almost nothing without the split. Here is the same 58 broken into pillars, which turns a vague grade into an ordered work queue.

Eighteen points sit in content and thirteen in schema, so those two pillars hold three quarters of the available upside. The technical gap of four is the tempting one — it is usually a couple of config changes — but it is capped, and spending a sprint there buys at most four points. The correct order is content first, because it is both the largest gap and the heaviest weight, then schema, where the FAQPage and Product line items alone can close most of a thirteen-point hole in a day of templating work.

The trap in this table is treating the gap column as a to-do list to be exhausted. Authority and freshness includes checks you cannot ship on demand — referring domains, mentions in high-trust communities, press coverage. Those seven points are real but they move on someone else's timeline, so plan them as a programme rather than a task. A site that closes content and schema alone lands in the mid-eighties without touching a single external signal.

• The score-chasing loop. A team optimises for the number rather than the behaviour behind it, stuffing thin FAQ blocks onto pages that have no questions and splitting coherent explanations into fake standalone paragraphs. The checks pass and the score rises while the page gets worse to read, which eventually shows up as citations that never arrive. The fix is to treat every point as a hypothesis about a real reader, and to reject any change you would not ship if the score did not exist.
• The homepage-only audit. Someone scans the marketing homepage, sees a respectable grade, and files AI readiness as done. But schema and extractability are template-level properties, and the homepage is usually the one template with the most attention and the least text. The fix is to score one URL per template — product page, docs page, blog post, category listing — because the weakest template is the one setting your actual ceiling across most of the site.
• The plateau at 84. A team closes every fix inside content and schema, lands in the low eighties, and concludes the score is broken because it will not move further. It is not broken: the remaining points are almost entirely in authority and freshness, which depends on referring domains, community mentions and press that no amount of on-site work produces. The fix is to reclassify that last stretch as a separate programme with its own timeline instead of grinding at pages that already pass.

Read top to bottom, that table is a zoom level rather than four alternatives. An answer capsule is one paragraph; citation density is how many capsules a page carries; the AI Score aggregates density alongside forty-six other properties into a site-level grade; GEO is the practice the whole stack exists to serve. The consequence is that you should never compare the score against citation share and expect them to track in lockstep. One is a snapshot of what you have built, the other a delayed report on how four independent retrieval layers reacted to it.