Key Takeaways
- Treat retrieval as a separate step from being known. An assistant can know your industry perfectly and still never fetch your page. Those are two different systems, and only one of them is yours to influence.
- Assume a handful of pages get read, not a page of results. The set that reaches the answer is small. Ranking tenth for your best question is closer to invisible than most people expect.
- Make one page unmistakably about one question. Pages that cover five things get passed over for pages that cover one. Breadth helps a human browse and hurts you here.
- Check that a machine can reach the page at all. A blocked crawler, a login wall, or a PDF behind a form removes you from the running before any judgment about quality happens.
- Judge yourself on the shortlist, not on being indexed. Being present somewhere is the floor. The only question that matters is whether your page is among the few that get opened for the question you care about.
There is a step between your buyer typing a question and an assistant naming three businesses, and almost nobody who is trying to get named knows it exists. It is a short step. It is also the only part of the process your website participates in.
That step is retrieval. The assistant goes out, fetches a few documents, reads them, and writes its answer from what came back. Everything you publish is competing for one of those few slots and nothing else.
Once you can see the step clearly, a lot of confusing behavior stops being confusing. Why a business with a beautiful site never gets mentioned. Why an assistant knows your industry cold and has never heard of you. Why the same question asked twice can return different names.
What is retrieval in AI answers?
Retrieval is the step where an assistant fetches a small set of documents before it writes. It turns your buyer's question into one or more searches, opens a handful of the results, reads them, and composes the answer out of what it found.
The word is worth knowing once and then setting aside. What matters is the shape of it: a search, a shortlist, a read, an answer. Your page has to survive all four to be in the room when the names get chosen.
Not every answer works this way. Ask what a business coach does and an assistant will answer straight from what it absorbed during training. Ask which business coaches in Denver work with first-time founders and it will usually go and look, because that answer depends on facts it has no reliable stored version of. The distinction is covered in more depth in what AI search is and how it differs from ordinary search.
Why being findable is not the same as being read
Being findable puts you in the queue. Being read means the assistant opened your page, and only a few pages get opened for any given question.
This is the gap that catches people out. A business checks that its page is indexed, sees that it is, and concludes the visibility problem must be elsewhere. But a search that returns your page in position eleven has technically found you and practically excluded you, because nothing at position eleven was going to be opened.
The pages actually opened and read
Read first
The page is plainly about the exact thing that was asked, not about the topic in general.
Read second
It answers in the first two sentences, so a passage can be lifted without the page around it.
Read third
It is text on an ordinary page: nothing to download, nothing to log into, no form in the way.
Read fourth
Something outside the site says the same thing, so the claim is not resting on one voice.
Everything else the search turned up. Present, permitted, indexed, and never opened. The queue length and the size of the read window here are illustrative rather than measured, and both vary by assistant and by question. The point they carry is the one stated in the section above and in the table below: the set that reaches the answer is small, and being in the queue is not being in it.
The useful reframing is to stop asking whether you can be found and start asking whether you would be one of the few opened. That is a much harder test and a much more honest one.
What retrieval does that stored knowledge cannot
Retrieval supplies the facts that are recent, specific, or checkable. Stored knowledge supplies the general picture. When an answer needs your current price, your new certification, or the fact that you exist at all, it needs something fetched.
This is why a model can be fluent about your entire field and blank about you. Your field is well documented in the material it was trained on. You are a small number of pages that may or may not have been in that material, and may or may not get fetched today.
It is also why the fix is not waiting for the next model. A new model has a newer stored picture, but the part of the system that could name you tomorrow is the part that goes and looks, and that part is already running. Perplexity's help center describes a product built almost entirely around this step, and its behavior is a good illustration of how much of an answer can come from material fetched at the moment of asking. The mechanics of how it picks are in how Perplexity chooses its sources.
What decides which pages get pulled in
Five conditions decide it, and every one of them is yours to change. They are ordered the way they get applied, which matters, because failing an early one makes the later ones irrelevant.
Most businesses that feel invisible are failing one of the first two, which are the least interesting and the easiest to fix. Very few are failing the fifth, which is the one everybody worries about.
| Condition | What it means | How to check it | What failing looks like |
|---|---|---|---|
| Reachable | A crawler is permitted to fetch the page and actually can | Open your robots.txt and read it; load the page in a logged out browser | Nothing about you comes back, even when asked by name |
| Readable | The words exist as text, not inside an image, a video, or a download | Try to select the sentence with your cursor. If you cannot, neither can a machine | Your strongest material is never the thing that gets quoted |
| Findable | The page surfaces for the words a buyer would actually type | Search your buyer's phrasing, not your own, and see whether you appear | You get read for a question nobody is asking |
| Matched | The page is obviously about that one question, not about the topic broadly | Read your title and first paragraph and name the single question they answer | A narrower page from someone else gets opened instead of yours |
| Quotable | One passage stands alone and still makes sense without the page around it | Read your best sentence out loud with no setup and see if it survives | You get opened and read, and then nothing of yours is repeated |
What this changes about what you publish
It changes the unit. Stop thinking in websites and start thinking in pages, because retrieval happens one page at a time and your site as a whole is never what gets fetched.
In practice that means one page per question you want to be the answer to. A single services page mentioning six things you do competes badly against six pages each doing one job, not because six is a magic number but because a narrow page matches a narrow question and a broad page matches nothing in particular.
It also means the boring technical layer earns its keep. OpenAI publishes documentation for the crawlers behind ChatGPT, including how to allow or block them, and Google's description of how its AI features in Search work is explicit that these answers are assembled per question from material that can be retrieved. A page nobody is allowed to fetch is not a weak candidate. It is not a candidate.
One honest limit. Doing all of this well makes you retrievable, and retrievable is not the same as recommended. Plenty of pages get opened, read, and then quietly left out of the answer because something about the business did not add up. Retrieval is the door, not the room. What happens after the door is the subject of what it means when AI cites a website.
Give changes time to show up, too. A page published this week may be fetched tomorrow or in a month depending on the assistant, so a re-check three days later tells you almost nothing.
If you would rather see the result than reason about the mechanism, the free check asks 25 real buyer questions from your niche through ChatGPT at $0 and emails a plain-English readout within the hour. It runs as soon as you ask for it.
Questions we hear the most
What is retrieval in AI answers?
Retrieval is the step where an assistant goes and fetches a small set of documents before it writes anything. It runs a search, opens a handful of results, reads them, and builds the answer out of what it found.
Does every AI answer involve retrieval?
No. Some answers come entirely from what the model absorbed during training, with nothing fetched at all. Assistants generally reach for live sources when a question looks current, local, or specific to a named business.
How many sources does an assistant actually read?
Far fewer than a page of search results. Published behavior varies by product and question, but the working assumption should be a handful of documents rather than dozens, which is why position in the shortlist matters so much.
If my page ranks well in Google, will it be retrieved?
Often, but not reliably. Search ranking is one input among several, and assistants apply their own judgments about whether a page answers the specific question asked, which is a narrower test than ranking for a keyword.
Can I make my site easier to retrieve?
Yes, and most of it is unglamorous. Let crawlers reach the page, publish the words as text rather than inside images or downloads, and give one page one clear job so it matches one question exactly.
How do I find out whether AI is retrieving my pages?
Ask an assistant your buyers questions in a fresh logged out session and look at what it cites. Your own wording coming back, or your domain appearing in the source list, is the clearest evidence that something of yours was fetched and read.
What does the AnswerHalo free check tell me about this?
It asks ChatGPT 25 real buyer questions from your niche and records who gets named and what gets cited. That shows whether your material is reaching the answers at all, which is the first thing to know. It costs $0.
Is the free check something I can run today?
Yes. It runs as soon as you ask for it, and the report is in your inbox within the hour, 7am to 5pm Central, or by 10am the next morning. The checks in this post can also be done by hand in an afternoon.
Your buyers are already asking. Find out what AI tells them.
The free check asks ChatGPT 25 real buyer questions about your niche and sends you the report within the hour, at $0. If it shows you are already getting named everywhere, we will say so in plain words.
Get my free check →