AEO Competitor Comparison: AI Visibility Side-by-Side

These AI-detected topics appear prominently on the competitor page but are absent from yours. Adding FAQ answers covering these topics could shift citation probability in your favour.

Every number in this report falls into one of two categories, and conflating them is the most common way teams misread an AEO audit. The first category is observable fact. Either ClaudeBot receives an HTTP 200 from your origin or it receives a 403. Either an llms.txt file exists at the root or it does not. Either the FAQ answers are present in the served HTML or they are assembled by JavaScript after load. These are binary, verifiable, and reproducible by anyone with curl.

The second category is inference. A "citation probability" of 87% for ChatGPT is a model built on the first category — it is a weighted expression of how many of the preconditions for retrieval your page satisfies relative to a comparison set. It is not a measurement of how often ChatGPT actually cites you, because no vendor has access to the retrieval logic inside any assistant, and none of the assistant operators publishes one. Nobody outside OpenAI, Anthropic, Google, Microsoft or Perplexity can tell you what determines a citation.

This distinction is what makes the comparison useful rather than decorative. A 14-point score gap is a summary. The ClaudeBot 403 is a fact, it is actionable this afternoon, and it is the single line in this report that would change reality if you fixed it. Read the facts first and the scores second.

A 403 to a named crawler is not a ranking problem or a content problem. It is an access problem, and it means one assistant's retrieval path to your page does not exist. The competitor returns 200 to the same request. Everything else in this report is a matter of degree; this one is a matter of presence or absence.

The cause is almost never deliberate. It is a WAF rule, a bot-management ruleset, or a rate-limiting policy that matches on a substring of the user-agent — commonly the literal token bot, or a vendor name such as anthropic — and drops or challenges the request before it ever reaches the application. Nobody decided to block Claude. Somebody enabled a managed bot-protection ruleset, and the default behaviour of that ruleset is to treat unfamiliar automated clients as hostile.

Diagnose it in the right order. Check robots.txt first, because a User-agent disallow is the one cause you can fix in thirty seconds. If robots.txt is clean, request the page as the crawler and observe what comes back — a 403 with a challenge page, a 429, or a connection reset all point at different layers. If the request never appears in your application access logs at all, the block is upstream of the origin, which means a CDN or WAF rule and not your web server config.

One caution about the word "reclaim". Restoring crawler access restores the possibility of retrieval. It does not restore a lost ranking, because there is no ranking to lose: assistants are not obliged to cite you and there is no queue you rejoin. What you get back is eligibility, which is a precondition and not a promise.

The lesson from this table is that "allow AI bots" and "block AI bots" are both incoherent instructions. The question is always which agent, for which purpose, and what you are trading away.

The signal table shows the competitor carrying FAQPage schema with seven question-and-answer pairs while your page carries none. It is worth being precise about what that difference does and does not do, because the usual framing is wrong.

It does not win them a rich result. Google restricted FAQ rich results to recognised government and health sites in 2023, and removed HowTo rich results entirely — which is also why the HowTo schema on your page, listed above as a strength, no longer produces anything in Search. It remains valid markup and it costs nothing to keep, but it is inert as a SERP enhancement and should not be counted as an advantage.

What FAQPage markup does buy is unambiguous machine-readable question-and-answer structure in the served HTML. An assistant summarising a features page has to work out which sentence answers which question. Explicit Question and acceptedAnswer nodes remove that guesswork. There is no verified evidence that assistants preferentially cite pages carrying more schema — the honest claim is narrower, that retrieval systems parse rather than render, so markup in the initial HTML response reaches them and disambiguates what they read.

The corollary matters more than the markup: the schema is worthless if the answers are not there. A page whose FAQ answers exist only inside JSON-LD, or only inside a collapsed accordion rendered client-side, has given a parser structure without substance. Write the answers as visible prose that a human would accept as complete, then mirror them into the schema. That order is not optional — structured data must describe what is on the page.

The four topics flagged above are absent from your page and prominent on the competitor's. The instinct is to add the words. That is keyword-density thinking, and density is not a ranking factor for Google and has no known role in assistant retrieval either. Mentioning "SOC 2 compliance" four times without holding SOC 2 certification produces a page that is both useless and untrue.

The correct reading is that each of these topics is a question a buyer asked somebody, and the competitor answered it on the page while you did not. "AI-powered task automation" is not a keyword, it is a capability a prospect wants to confirm before booking a demo. "Project management API" is a developer checking whether integration is possible at all. "Free forever plan" is a pricing objection being resolved before the pricing page is ever loaded.

So the action is: for each topic you can honestly claim, write a direct, self-contained answer in visible prose — the kind that stands alone if lifted out of the page — and mirror it into the FAQPage schema. For each topic you cannot honestly claim, do nothing. A gap that reflects reality is not a defect, and closing it with words rather than facts is how a features page stops being trustworthy.