Framework
The four layers of AI answer failure
"We don't show up in ChatGPT" is four different problems wearing the same sentence. They have different causes, different tests, and different fixes. Working on the wrong one is the most common way companies waste a year on this.
An engine does four things between a buyer's question and the answer on screen. It reaches your pages. It decides whether to use them as a source. It forms an understanding of what your company is. And it writes a description of you into the answer. Each step can fail independently, and each one can fail while the step before it worked perfectly.
Most tools in this category measure the first two and report a percentage. The failures that cost pipeline tend to live in the last two — where a model reads your page, takes a fact from it, recommends a competitor, and describes you as the company you stopped being.
Why the order matters
The layers are sequential. Content published on top of an unresolved identity does not compound — it gets filed under whichever record the model already settled on, and adds weight to the wrong one. If the model thinks you are a chatbot company, ten excellent pages about AI agents become ten excellent pages by a chatbot company.
So the diagnosis runs downward and the work runs upward. You find the lowest layer that is broken, and you fix from there.
1 Retrieval
Can the engine reach and read your pages at all?
How it fails
- The page renders client-side and the crawler receives an empty shell. Most AI crawlers do not execute JavaScript.
- robots.txt allows the bot, but a WAF or bot-fight rule returns 403 to the same user agent.
- The server returns less content to a bot than to a browser — not a block, a quiet downgrade.
- The page exists but is in no sitemap and nothing links to it.
How to measure it
Request the page as each bot and compare the response, word count and status code against what a browser receives. A 200 is not enough; the body has to match.
What resolves it
Server-side render or pre-render the pages that matter. Allow search and user-triggered bots at the firewall, not just in robots.txt. Verify bots by IP range instead of blocking the user agent.
2 Citation
Does the engine use your content as a source?
How it fails
- The page is retrievable but answers nothing a buyer asks, so nothing pulls it into an answer.
- A competitor covers the same question with more sources pointing at it.
- The content is correct but undated, so the engine prefers a fresher-looking source.
- The claim worth citing is buried in the middle of a long page instead of stated plainly.
How to measure it
Run the buyer question and record which URLs the engine cites. Yours either appears or it does not. Per engine, because they disagree.
What resolves it
Write the page that answers the question directly, state the citable claim early, give it a visible date, and earn mentions on pages you do not control.
3 Interpretation
Does the engine understand what your content says?
How it fails
- The page says what you do but never names the category, so the model infers one.
- Your title, meta description, H1, schema and llms.txt each describe the company differently. The model picks one.
- Two live domains split the citations, and neither accumulates enough to settle the record.
- Wikidata, a directory profile, or an old press release carries a category you abandoned — and it has more sources behind it than your homepage.
How to measure it
Compare every field on your own site against each other first. Then compare the consensus of your site against what third-party sources say. Disagreement is the finding.
What resolves it
One description of the company, repeated identically everywhere you control. Then correct the third-party records, starting with the ones that have the most citations pointing at them.
4 Expression
How does the engine describe you in the final answer?
How it fails
- You are cited accurately and described in the category you left two years ago.
- You are mentioned in a list of ten, which is not the same as being recommended.
- The description is accurate but aimed at a buyer you do not sell to.
- The engine names a competitor as the answer and uses your page as the source for why.
How to measure it
Ask the question your buyer would ask, with their constraints attached, five times per engine. Record who was recommended, in what order, and how you were described when you appeared.
What resolves it
This layer is downstream. It moves when the first three are resolved and the pages that decide the answer exist — comparisons, alternatives, integrations, "best for" pages built around constrained questions.
Using it
Start at layer one, because it is the cheapest to test and the most common silent failure. Request your own page as a bot and compare it to what a browser gets. If the bot receives an empty shell or a 403, stop — nothing above that layer can work, and no amount of content will change it.
If retrieval is clean, check whether your own site agrees with itself before looking anywhere else. Title, description, H1, schema, llms.txt. Internal disagreement is free to fix and it is where a surprising number of these cases end.
Only then is it worth running buyer questions across engines. By that point, a bad answer tells you something real: the record describing you is wrong, or it is right and the pages that decide the answer do not exist yet.
What this framework does not cover
It describes failure modes, not timelines. Changes to your own pages register soonest in engines that search the web on every question. Changes to how third parties describe you depend on those sites updating and being re-crawled, which is outside anyone's control. Any framework that promises a date for a specific engine is guessing.
It also does not cover whether the buyer wanted your product in the first place. Being described correctly is table stakes, not a strategy.