Can AI Answer Real Questions About You? The Answer Pillar Explained

Here is a test almost everyone runs, and almost everyone reads wrong.

You open ChatGPT, type your company name, and ask what it does. The answer comes back broadly right. You conclude your site is doing its job, close the tab, and move on.

That test is rigged in your favour, and it is the reason the fourth pillar exists.

When you ask an AI about your business by name, it can answer from anywhere. Its training data. Your Google Business Profile. A directory listing. A review site. An old press mention. Your competitor's comparison page. Any of those can produce a correct answer while your own website sits there saying almost nothing useful about you.

You have not tested your site. You have tested the internet's collective memory of your business, and then given your site the credit.

Answer is the fourth and final pillar of an AI Readiness score, and it fixes that by closing the book.

The Closed-Book Test

The pillar works by asking an AI model six fixed questions about your business, with one constraint that changes everything: it can only use what the crawler found on your site. No training data. No search. No directories. No Google Business Profile. Just your pages.

That constraint is the whole design. An open-book test tells you whether the world knows who you are, which is interesting and largely outside your control. A closed-book test tells you whether your site says who you are, which is entirely inside your control and is the only part you can actually fix.

It also explains why so many businesses are surprised by the result. A company can be well known, well reviewed, and thoroughly described across the web, and still have a website that, read on its own, cannot answer where the business is located.

The same business, two tests, two opposite results

Take a composite example, assembled from patterns we see repeatedly rather than one real client: a mid-sized accountancy firm, established, well reviewed, listed everywhere, twelve years in business.

Open book. Ask an assistant "what does Hartley & Fern do and where are they based?" and you get a fluent, accurate paragraph: chartered accountants, Manchester, specialising in owner-managed businesses. Every word correct. The firm's marketing lead sees this, concludes the site is doing fine, and stops looking.

Closed book. Give a model nothing but the firm's own homepage HTML and ask the same thing. It can tell you they are accountants, because the page says so in a heading. It cannot tell you where they are, because the address is inside the footer logo image. It cannot tell you what they specialise in, because that lives in a slide carousel that only populates after JavaScript runs. It cannot tell you who runs the firm, because the partners are on a separate page that the homepage links to as "Meet the team."

The open-book answer was correct because the internet remembered. A Google Business Profile, a directory listing, twelve years of local coverage. None of that is the site, and none of it is under the firm's control the way their own HTML is.

The gap between those two results is exactly what this pillar measures. A large gap means you are being carried by external sources, which works right up until an engine leans more heavily on first-party data, or a directory listing goes stale, or a competitor's page becomes the better match.

The Six Questions

Six fixed questions, the ones a real customer would need answered before doing business with you:

• What does this company do?
• Where is it located?
• What areas or industries does it serve?
• What specific services or products does it offer?
• Who runs it?
• How can someone contact it?

There is nothing clever about the list, and that is deliberate. These are not trick questions or edge cases. They are the minimum a stranger needs, and a site that cannot answer them from its own content has a problem that no amount of content marketing fixes.

Each answer is then checked for one further thing: whether the fact traces back to something your site actually declared, or whether the model reconstructed it from prose and happened to be right. That is entity grounding, and it is worth understanding on its own, because a correct guess and a grounded fact behave completely differently the next time anything changes.

What Failure Actually Looks Like

The output that lands hardest is not a score. It is a sentence with two halves.

Asked what your company does, an AI answered: "a WordPress plugin developer." Your site says: an AI visibility platform.

Nothing is broken. No validator failed. No page returned an error. A machine read the site carefully, formed a confident and specific impression, and got the business wrong.

This is the failure mode that makes the pillar worth having, because it is invisible from every other angle. Your analytics will not show it. Your rankings will not show it. The only way to find it is to ask, and then to compare the answer against what you know to be true.

The second most common shape is worse in a quieter way: the model cannot answer at all. Asked where the business is located, it has nothing. Not a wrong answer, an absent one, on a site whose address is sitting in an image in the footer where no crawler will ever read it.

The third shape is the sneakiest, because it looks like success. The answer is right but far too general. Asked what the company does, the model says "a professional services firm," which is true, useless, and indistinguishable from ten thousand other businesses. A site that only supports a generic answer will only ever produce a generic mention, and generic mentions do not win customers.

Why It Is Only Worth 25 Points

Answer is the pillar closest to the question people actually care about, so it is fair to ask why it carries 25 of the 100 points rather than the largest share.

The reason is that it is a symptom measure, not a cause measure. When Answer scores badly, the fix is almost never in Answer. It is upstream. The crawler could not reach the pages, so there was nothing to read. The pages were unreadable, so the content did not survive. The entity data was missing or disconnected, so the facts were never declared. Fix the graph and Answer tends to improve on its own.

Weighting it higher would mean paying twice for the same problem: once where it originates and again where it shows up. Twenty-five points is enough to matter and not so much that a single upstream failure gets counted three times.

It is also the pillar with the most inherent variability. Language models are not deterministic, and an answer can differ slightly between runs. That is a real limitation of this kind of measurement, and it is a reason to treat the pillar as directional rather than precise, which the weighting reflects honestly.

The Page-Level Version

Underneath the six questions, a signed-in audit also runs a per-page comprehension pass: it goes through your pages and lists which ones a machine could not make sense of.

That list is deliberately not scored. It is a work queue. Scoring it would invite the wrong behaviour, which is optimising a number rather than fixing the pages. A plain list of URLs that failed to parse is more useful and harder to game.

The anonymous free audit runs a shallow two-question version of the main test as a preview. The full six-question scored run and the per-page pass need a signed-in audit, mostly because a full crawl is expensive rather than because we are hiding the interesting part.

What This Pillar Refuses to Claim

Worth stating plainly, because it is the line between a useful measurement and a fabricated one.

It does not measure what ChatGPT thinks of your business. It measures what an AI concludes when given only your site. Those are different questions, and only the second is something you control or we can honestly test.

It is not a prediction that you will be cited. No part of this forecasts whether an AI will mention you, and anyone attaching a percentage to that is quoting a number nobody can substantiate.

It never invents a verdict about the live engines. Where the product offers to check what the real engines say, it opens ChatGPT, Perplexity, Gemini or Google's AI mode with one of these same six questions pre-filled, and you read the answer yourself. That is framed as "go and look," not as a measurement we performed, because we did not perform it.

The distinction running through all of this is readiness versus visibility: what your site makes possible, versus what the engines choose to do. This pillar sits as close to the boundary as anything can while still measuring your side of it.

Common Mistakes

Running the open-book test and believing it. Asking ChatGPT about your business by name is a fine thing to do and a bad way to judge your site, for the reason at the top of this post. If you want the flattering number, ask by name. If you want the useful one, take the question to your site.

Fixing Answer directly. Stuffing a paragraph of identity facts into your footer because the test said your location was unclear treats the symptom. The location was unclear because it was never declared. Declare it.

Assuming a right answer means a good answer. A correct guess is not a grounded fact. It survives until your homepage copy changes or the model updates, and then it does not.

Reading a single run as gospel. Models vary between runs. A one-point movement is noise. A question that fails consistently across several runs is a finding.

Treating an absent answer as less serious than a wrong one. In practice, absence usually means the fact exists only in an image, a PDF, or JavaScript that never ran. That is a plumbing problem with a specific fix, and it is often the fastest win on the whole list.

Only ever testing the homepage. Your homepage is the most carefully written page on your site and the least representative. The pages a customer actually lands on from a search are usually thinner.

Forgetting that customers run this test too. The six questions are not an abstraction. They are what someone asks before deciding whether to contact you, and increasingly they ask an assistant rather than reading your homepage.

How to Run a Closed-Book Test Yourself

You can approximate this in about fifteen minutes without any tool.

1. Get your page as a machine sees it. Open your homepage, right click, choose "View Page Source." That is the HTML your server actually sent, before any JavaScript ran. Copy it.

2. Paste it into an AI assistant with a hard instruction. Something like: "Using only the page source below and nothing you already know, answer these six questions. If the page does not say, answer 'not stated.'" Then paste the six questions and the source.

3. The "not stated" answers are your list. Those are the facts your site is currently leaving to guesswork, luck, or someone else's directory listing.

4. For each wrong or missing answer, find where the fact lives. Usually it is in an image, a PDF, a contact form, or nowhere at all. Occasionally it is in prose the model read differently than you expected, which is its own useful finding.

5. Repeat on your most important non-homepage page. Homepages are usually the best-described page on a site. A service page or product page is a fairer test of the rest.

6. Fix upstream, then re-run. Declare the facts properly rather than writing them into a paragraph aimed at the test. Then run it again and see whether the "not stated" list shrank.

If your site fails badly on this, the causes are usually the same handful, and we went through them in why ChatGPT often cannot answer basic questions about a real, established business.

The Four Pillars, Complete

That closes the series.

Discover asked whether a crawler can reach you at all. Fifteen points, the fastest to fix, and invisible from a browser.

Read asked whether the page can be parsed once the crawler arrives. Another fifteen, mostly structural.

Understand asked whether the site says who you are and connects its pieces together. Forty-five points, the heaviest, and the one where most sites with "good schema" quietly lose.

Answer asks whether all of that adds up to a machine being able to describe your business correctly. Twenty-five points, and the one that tells you whether the other three worked.

They run in that order for a reason. A failure at any stage makes everything after it meaningless, which is why a bad Answer score is almost always a message about something further up.

Curious what an AI concludes about your business when it only has your website to go on? Run a free AI readiness check and see the answers, alongside what your site actually declares.