AI Extractability Benchmark: 200 Pages Scored for Answer-Engine Citation
Ranking and citability are different questions about the same page. This is the full, per-page record of that second question for medicaidworknavigator.com — every URL in the sitemap, scored on five on-page signals, published as CSV and JSON so the method can be checked or reused.
Quick answer: across 196 content pages the median score is 95/100 and the lowest is 40. 184 of 196 content pages state a figure in their own summary, and 12 do not — that is the single largest gap in the set. Snapshot taken 2026-10-19.
How the score is built
Each page gets a 0–100 index over five signals, weighted as below. The crawler reads the live HTML of every sitemap URL; nothing is inferred from traffic, and search data is used only to order the table.
| Signal | Weight | What it measures |
|---|---|---|
| Structured data | 30 | Article, FAQPage or a tool type declared in JSON-LD. Without it an answer engine has to guess what the page is about. |
| Q&A density | 25 | How many questions the page answers in Q&A form — the unit an extractive engine lifts most often. |
| Figure in the summary | 20 | Whether the page's own description states a number. For “how much can I earn” queries, the number in the snippet is the answer. |
| Declared date | 15 | A visible date-modified — benefits rules change annually, and undated pages age badly. |
| Threshold tables | 10 | Income and hour thresholds in a table are the passage most often quoted whole. |
One signal — the declared date — scored identically on every page in this snapshot, so it adds nothing to the ranking. It is reported rather than hidden, because a signal that cannot discriminate is not evidence of quality.
Score distribution
The distribution is heavily top-loaded: 78% of content pages score 90 or above. That is itself the finding — a site can look uniform and still lose the queries where no page puts a number in its summary.
Every page, worst first
196 content pages. Utility pages (privacy, terms, editorial policy, about) are in the CSV but excluded here — they carry no editorial content to benchmark.
What the two gaps mean
16 content pages score under 80
Most are missing one specific block, not the whole structure. The lowest — /, /states/south-carolina/, /states/alabama/ — are the pages to fix first if the goal is being quotable rather than merely ranked.
12 content pages put no figure in their own summary
This is the signal with the clearest cost: for a “how much can I earn” query, the description is often the whole answer an engine has to work with. It also includes several pages that already earn impressions, so the missing number is visible on the search result itself.
Reading this dataset
What is an AI-extractability score?＋
A 0–100 index over five on-page signals — structured data (30), Q&A density (25), whether the summary states a figure (20), a declared date (15) and threshold tables (10). It measures whether the page contains something an answer engine can lift verbatim. It is not a citation count and it does not predict ranking.
Can I compare these scores with another website?＋
No. The index is relative to the pages in this snapshot. Two sites scored by the same crawler are not on a common scale, because the weighting is a checklist heuristic rather than a calibrated measurement. Compare a page with its own site, over time.
Why are 200 pages published if only 196 have content?＋
Four utility pages — privacy policy, terms, editorial policy and the author page — were crawled because they are in the sitemap. Their scores are published for completeness, but they are excluded from the distribution and the benchmark table because they carry no editorial content to extract.
What does a perfect score mean?＋
That the page declares structured data, answers questions in Q&A form, states a figure in its own description, carries a visible date and includes a threshold table. It means the material is present and machine-readable — not that an engine will choose it over a .gov source.
How often is this refreshed?＋
The snapshot is dated at the top of the page. Re-running the crawl produces a new dated file; older snapshots are not overwritten in prose, so a score can always be traced to the date it was measured.
How should I cite it?＋
The citation block at the bottom of this page carries the author, year, title, publisher and URL, under CC BY 4.0. The CSV and JSON carry the same citation in their metadata.
Cite this dataset
Researchers and writers may use this data with attribution. It is published under CC BY 4.0 with a stable URL, so a score can be cited and re-checked.
Scope and limits: this benchmark scores pages of one site against each other. It is not a citation count, not a ranking factor, and not comparable across sites. Whether an AI engine actually quotes a Medicaid page also depends on the engine's own source preferences — for benefit-rule queries those lean toward .gov sources, which no amount of on-page structure changes. Snapshot: 2026-10-19. Last reviewed: 2026-10-19.
By Alyssa Xu, Founder & Editor.
Reviewed by Gavin YE, Technical Director, ClearRules Labs.
Page last reviewed 2026-10-19