20 checks that decide whether ChatGPT, Perplexity and Google AI recommend you
20 checks, 3 minutes → your score and the five fixes worth doing first.
Answer engine score
0
out of 100, across 20 checks
Not started
Answer the checks below. The score updates as you go, and the fix list reorders itself around whatever you are missing.
3 minutes • no signup • no crawl • nothing leaves your browser
What you get
- Identify the single unfixed check worth the most points, ranked automatically.
- See your score split across five weighted dimensions, not one flat number.
- Get a fix list sequenced by value, gap and effort — not by what is easiest to sell.
- Spot the three claims other checklists oversell, and why we scored them low.
By dimension
The order to do it in
Your fix list
Ranked by what each check is worth, multiplied by how much of it you are missing, divided by what it costs you to fix. Not by what is easiest to sell.
Answer a few checks and the sequence appears here, highest-leverage first.
Evidence grading
What we actually know, and what is being oversold
Most AEO checklists present folklore and finding at identical confidence. Every check in this scorecard carries a grade, and three of them are weighted down on purpose.
Corroborated by primary research or large-sample independent analysis. The GEO paper (KDD 2024) found adding statistics, quotations and cited sources measurably raised visibility in generated answers. Large-sample crawl analysis found AI-cited content runs materially fresher than typical organic results. Blocking AI crawlers demonstrably removes you from retrieval.
The mechanism is plausible and practitioners agree, but no controlled study isolates the variable. Question-shaped headings matching query fan-out, answer-first paragraphs, entity consistency, community presence. Sensible to do. Not proven in the way the first group is.
Widely repeated, thinly supported, and in one case actively contradicted. Schema markup as an AI citation lever has no vendor confirmation and Google has been removing schema features, not adding them. We score it, lightly, and say so.
Three claims we deliberately left out
- llms.txt — proposed September 2024 to help coding assistants find clean documentation, not as a ranking signal. Google has stated publicly that it does not use or endorse it, and crawler monitoring shows AI agents almost never request the file. Publish one if you like; it is not a visibility lever and does not belong on a scorecard.
- FAQPage and HowTo schema as AI levers — FAQ rich results were restricted to authoritative government and health sites in August 2023 and fully deprecated in May 2026. HowTo was removed in 2023. The question-shaped content underneath the markup still matters. The markup is not what is doing the work.
- "Longer, more comprehensive content wins in AI answers" — large-sample analysis found near-zero correlation between content length and citation probability, with the majority of AI Overview citations going to pages under 1,000 words. Length is not the variable. Extractability is.
This is a self-assessment, not a crawl. It reads nothing about your site — you supply the answers, which means it is exactly as honest as you are. That is deliberate: the automated checkers that do crawl can only see four of these twenty things, and the other sixteen are where the real gap usually is.
Questions people ask about AI search visibility
How do I get cited by ChatGPT and Google AI Overviews?
Five things have to be true, in order. The crawler has to be allowed in. The content has to exist in the HTML without running JavaScript. A self-contained passage has to answer the question inside the first 80 words. The claim needs evidence a model has reason to prefer — original numbers, named sources, a named author. And somebody other than you has to say the same thing. Most sites fail at step one or two and spend all their effort on step five.
Is AI search optimisation different from SEO?
It is a layer on top, not a replacement. Classic ranking signals still decide which pages enter the pool that gets retrieved — so authority, links and technical health remain load-bearing. What changed is selection within that pool, which is increasingly passage-level rather than page-level. One large-sample analysis found the overlap between AI Overview citations and the traditional top ten fell from 76% to 38% in about seven months. Ranking is necessary and no longer sufficient.
If AI answers the question, does the traffic still matter?
Less of it arrives, and the part that does is worth more. US searches ending without a click reached 68% in early 2026, up from around 60% in 2024. Pew found users clicked a result 8% of the time when an AI summary appeared, against 15% when it did not. But the same research finds brands cited in AI answers earn substantially more clicks per impression than uncited ones. The pool shrank; being in the answer got more valuable, not less.
Why do interactive tools survive this better than articles?
Not because answer engines favour them — there is no evidence for that, and anyone telling you otherwise is selling something. It is simpler. An AI answer can replace a fact. It cannot replace an interaction. When the value of the page is a calculation performed on the reader's own numbers, summarising the page does not satisfy the query. The article that explains a formula is replaceable. The thing that runs the formula on your figures is not.
How do I measure whether any of this worked?
Ask the assistants. Write down the ten questions a buyer would ask before choosing someone in your category, run them monthly across ChatGPT, Google AI Mode, Perplexity and Claude, and record whether you appear, in what position, and described how. It is manual and it is unglamorous, and it is a better baseline than any dashboard, because you are measuring the actual output rather than a proxy for it.
Next
Knowing the gap is half of it. The other half is the questions.
Answer-first content only works if you are answering the questions people ask. The Question Map generates them for your market and tells you which page each one belongs on.