How to Score AEO Content Quality Without a Vanity Number

Score AEO content quality with a short review rubric, not a single number. Rate each draft on five lenses: usefulness, evidence, clarity, entity accuracy, and technical access. Give each a 0, 1, or 2 against written anchors, treat technical access as a pass/fail gate, and block publishing on any zero for usefulness, evidence, or entity accuracy. What you get is a profile that tells the editor what to fix. It doesn’t predict rankings or AI citations, and it shouldn’t pretend to.
Here’s the situation this is for. You run an agency, your team ships two or three articles a month, and someone bought a tool that stamps each draft “87/100 AEO-ready.” Nobody can say what the 87 means, what moved it, or whether the page actually answers anything.
That’s the vanity number problem. This post gives you the replacement, where each lens comes from, and one worked example scored end to end.
Why does a single AEO score mislead you?
A total hides the one failure that matters. A page can be well structured, fast, and full of FAQ schema, then lose its value because one statistic is made up. Add the points together and that page still looks like an 85.
There’s also no outside number to calibrate against. Google’s own guide to optimizing for generative AI features says plainly: “No third-party tool has access to our internal ranking or AI systems.” The same guide warns that indexing and serving aren’t guaranteed, even when a page meets every requirement.
So any tool selling a “citation likelihood” score is grading its own guess. Handy for workflow, maybe. Not a quality verdict.
The guide also says that “optimizing for generative AI search is optimizing for the search experience, and thus still SEO.” We agree, and the rubric below is built on that. For the longer version, see what answer engine optimization means for B2B.
Here’s the shift from a vanity score to a review profile.
- One number, like 87/100
- Weights hidden inside a tool
- A fake stat can still pass
- Implies ranking or citation odds
- Five ratings of 0, 1 or 2
- Written anchors anyone can check
- A zero on a core lens blocks publishing
- Judges the page, not the algorithm
What are the five lenses in the rubric?
We call it the Five-Lens Review. There’s nothing secret in it. Each lens maps to a question Google already publishes for site owners, which is why we trust it more than a tool’s hidden weighting.
1. Usefulness. Does the page answer one real question, early, for a specific reader, and add something beyond common knowledge? This comes from Google’s helpful content questions, such as whether the content provides “original information, reporting, research, or analysis,” and from the AI guide’s push for “non-commodity content.”
2. Evidence. Does every material claim link to a source that resolves today, and is first-hand experience labeled as first-hand? Google’s helpful content page asks whether content shows “first-hand expertise and a depth of knowledge.”
3. Clarity. Can a reader find the answer from the headings alone? Google’s AI guide recommends organizing pages into paragraphs and sections, with “headings that provide a clear structure.” Short paragraphs, plain words, one idea per section.
4. Entity accuracy. Are the names, prices, products, dates, and people correct and consistent everywhere they appear? That includes your brand name, your author byline, and your structured data. Google’s structured data policies say: “Don’t mark up content that is not visible to readers of the page.”
5. Technical access. Can a crawler fetch the page and read the answer as text? Google says a page must be “indexed and eligible to be shown in Google Search with a snippet” to appear in its generative AI features. This one’s a gate, not a grade.
Here’s the one question each lens asks.
- Usefulness: does it answer one real question early and add something new?
- Evidence: does every material claim link to a live source or labeled first-hand note?
- Clarity: can a reader find the answer from the headings alone?
- Entity accuracy: are names, prices, dates and schema correct and consistent?
- Technical access: is the answer indexable, in HTML text, and allowed in snippets?
How do you score each lens without guessing?
Write down what a 0, 1, and 2 look like before anyone scores anything. Anchors stop two reviewers from arguing about whether a page “feels” like a 7.
| Lens | 0 (blocks publish) | 1 (needs a named fix) | 2 (ready) |
|---|---|---|---|
| Usefulness | No clear answer, or restates what’s already everywhere | Answer exists but sits low on the page or lacks a point of view | Direct answer near the top, specific reader, original angle or worked example |
| Evidence | Any unsourced or unverifiable statistic, or an invented quote or result | Sources exist but some are stale, secondary, or don’t support the exact claim | Every material claim links to a live primary source; first-hand notes are labeled |
| Clarity | Wall of text, vague headings | Readable, but headings don’t match the questions people ask | Question-shaped headings, short paragraphs, table or visual only where it helps |
| Entity accuracy | Wrong price, product, date, or name | Inconsistent naming, or schema that drifts from the page | Names, figures and schema match the visible page and your other pages |
| Technical access | Fail: blocked, noindex, nosnippet, or answer only in an image or script | n/a | Pass: crawlable, indexable, answer in HTML text |
Three rules turn the table into a decision:
- Technical access is pass/fail. A brilliant page nobody can fetch scores nothing. Check your robots.txt for AI crawlers and your JSON-LD here.
- A zero on usefulness, evidence, or entity accuracy blocks publishing. You can’t average it away.
- A 1 needs a named fix and an owner. “Evidence 1: replace the 2023 survey link, Priya, by Friday” is a review. “Evidence 1” on its own is just another number.
Clarity is left off the blocking list on purpose. A clarity 0 is usually a heading and paragraph edit that doesn’t change what the page claims. So the editor fixes it in the same pass instead of sending the draft back.
Here’s the publish decision as a tree.
What does a scored page look like? (Hypothetical worked example)
The example below is hypothetical. It’s a composite built to show the method, not a client result.
A 14-person Shopify agency drafts “How much does a Shopify store migration cost?” Draft 1 opens with 300 words on the history of ecommerce. It says “73% of merchants overspend on migrations” with no source. The price table is a screenshot.
The footer says “Northside Digital” while the byline and Organization schema say “Northside Studio.” And the pricing tab only renders after a click, through JavaScript.
The reviewer scores it:
- Usefulness 1. There’s an answer, but it’s buried, and it’s generic.
- Evidence 0. The 73% figure has no source. A search turns up nothing, so it goes.
- Clarity 1. Headings like “Things to Consider” don’t match how anyone asks the question.
- Entity accuracy 0. Two brand names, and schema that doesn’t match the footer.
- Technical access: fail. The price ranges live in an image and a click-to-load tab.
Draft 2 opens with a two-sentence answer: the agency’s own quoted ranges for small, mid-size, and complex migrations, labeled “from our 2025-2026 proposals.” The stat is gone, and the table is HTML.
Headings become questions: “What drives the price up?” and “When is a migration not worth it?” One brand name appears everywhere, and the schema matches it.
Here’s how the profile moved. Technical access shows as 0 for fail and 2 for pass so it fits the grid.
Notice what a total would’ve said. Draft 1 sums to 2 out of 10, which sounds bad but doesn’t tell anyone where to start. The profile does: kill the stat, fix the brand name, move the prices into HTML text.
In this example, those three fixes unblock the page. The new opening and question headings then lift usefulness and clarity to 2.
And notice what didn’t move the score: word count, keyword density, FAQ schema count. Google’s helpful content page asks whether you’re writing to a particular word count because you’ve heard Google has a preferred one. Its answer: “No, we don’t.”
A first-hand note from Satya Phanindra Reddy, founder of Magic Teams AI. The drafting pipeline behind this blog uses a weighted rubric with a publish gate at 80, and it stores the list of failing criteria next to each score. In practice the number only decides whether a post ships. The list is what the editor works from. If you keep one output from your scoring, keep the list.
Who should do the scoring, and when?
Score twice. The writer scores before handing off, then an editor scores without seeing the writer’s numbers. Where they differ by two points on any lens, talk it through. That conversation is usually where the real problem shows up.
An AI assistant can do a first pass on clarity and flag claims that need sources. It can’t confirm that a link supports the exact claim, or that your price is current. A person has to click every evidence link and check every entity.
There’s a reason for that caution. In a March 2025 test, the Tow Center for Digital Journalism found that eight AI search tools collectively gave incorrect answers to more than 60 percent of its citation queries. Don’t hand the evidence check to the same kind of system you’re hoping will cite you.
Re-score when facts change. Pricing, legal, and product pages drift fastest, so give them a review date. If you want the step before scoring, the AEO content brief sets the query and evidence requirements the reviewer checks against. The AI search content system post shows where review sits between brief and publish.
A caveat for law, accounting, and advisory practices. On pages about legal, tax, or financial questions, “Evidence 2” should mean a qualified professional has reviewed the claims, not just that the links resolve. This rubric checks editorial quality. It isn’t legal or financial advice, and it doesn’t replace your professional obligations.
How is this different from an AEO content audit?
Timing and scope. The Five-Lens Review is a pre-publish gate for one draft, so it leans on clarity and evidence while the page can still change cheaply.
An AEO content audit works across pages you’ve already published and adds a links column, because a library’s internal linking only shows up at that scale. Use the review on every new draft. Run the audit once a quarter or so on the back catalog.
How do you measure results without turning them into the score?
Keep quality and outcomes in separate columns. The rubric judges the page you control. Outcomes tell you what the market did with it, and they move for reasons outside the page.
For Google, the Search Generative AI performance reports in Search Console show impressions, pages, countries, devices, and dates for generative AI features such as AI Overviews and AI Mode, plus a separate Discover view. Google’s note on that post says the reports reached all websites worldwide as of August 31, 2026. Pair them with referral sessions and enquiries.
Then read the two columns together over a quarter. If pages scoring all twos still get no impressions, question the topic before you rewrite the prose. If you want a money figure on the work, use the same discipline as measuring ROI on AI automation: estimate, label it as an estimate, and don’t claim attribution you can’t show.
FAQ
Should I publish my AEO score on the page?
No. Readers don’t care, and a visible score invites the gaming the rubric is meant to stop. Keep the profile in your content log with the review date.
Can I convert the five lenses into a percentage for reporting?
You can, but report the profile next to it. A percentage without the lens breakdown brings the vanity number right back.
Does a perfect profile guarantee AI citations?
No. Google states that indexing and serving aren’t guaranteed even when a page meets every requirement. The rubric raises the floor on quality. It can’t promise placement in any engine.
Do I need llms.txt or special markup to pass technical access?
Not for Google. Its AI optimization guide says Google Search doesn’t use llms.txt or similar AI-specific files, and that structured data isn’t required for generative AI search. Other services may treat them differently, so decide per platform.
If your team wants brief, score, and measurement connected in one reviewable workflow instead of three spreadsheets, that’s the kind of operating problem we’re happy to talk through on a short fit call.