Why I tried it
I entered the feeder insect business, which is not a new business. Companies have sold crickets and worms to reptile owners for decades, and the big names were already well established online. Winning a two or three word Google term against them looked unrealistic for a newcomer.
So I went somewhere else. I read the public documentation from the AI platforms and thought about what a system that answers questions would want from a website. Then I wrote for the long, specific, high-intent questions that people actually type or speak, plus the comparisons they make before they buy.
In about a month, the store went from 915 pages on Google to 4,089. That number is real and it is the easy part of the story. The harder part is what it did and did not buy us.
What the engines say they do
Start with the mechanics, because they explain why long questions matter. Google, OpenAI, Anthropic and others describe a similar process. The assistant takes a question and breaks it into several narrower searches, sometimes called query fan-out. It then retrieves pages and cites what it found.
Eligibility is plain. A page has to be reachable by the engine’s search crawler and indexed. OpenAI documents separate crawlers for search and for model training, and the search crawler is the one that matters for being cited in answers (OpenAI crawler documentation). Google states that it has no special requirements for appearing in its AI features beyond its normal search guidance (Google AI features documentation).
Microsoft’s Bing guidance adds practical habits. Give each URL one topic, put the answer first, state facts explicitly, and keep your entity names consistent. None of this is exotic. It is clear writing that a machine can lift a sentence from.
What scaled, specific pages did
First, they got indexed. A store with 915 pages can only be retrieved for 915 pages’ worth of questions. A store with 4,089 pages has more doors, and each door is built for one question.
Second, they matched the shape of the searches. When an assistant fans a question out into narrower queries, a page titled for a precise question has a real chance of matching one of them. A page about a broad category rarely does. This is the part of the strategy I would repeat without hesitation.
Third, they gave the AI something specific to cite. A sentence that directly answers “which feeder is better for a juvenile of this species, and why” is easy to quote. A paragraph of general brand copy is not. Specific pages gave the assistants clean material.
I call the underlying idea epistemic retrieval. That is my own term, not an industry standard. It describes how an AI goes looking for the source that actually knows the answer and then cites it.
The engines document the retrieval mechanics. They do not claim to measure “knowing,” so treat my label as a working theory.
What volume alone does not do
Here I want to be careful, because the 4,089 figure invites the wrong conclusion. Page count is not a ranking factor inside an AI answer. Several engines have not documented how they rank or choose among the pages they retrieve, and anyone who tells you otherwise is guessing.
Volume also does not make a thin page useful. Hundreds of near-identical pages that swap one keyword and repeat the same paragraph give an assistant nothing new to cite. They also give a human reader nothing to trust. Google says plainly that trust is the most important part of how it evaluates quality.
And volume does not replace proof. A page can answer a question perfectly and still not be the source the AI names. That brings me to the field test that changed how I think about this.
What the answers actually showed
In October I asked several assistants for local recommendations. The cities were Richmond, Virginia and DeLand, Florida. The categories were coffee, plumbers, air conditioning repair and pizza, and I ran Google AI Mode and ChatGPT logged out, with Claude and Gemini on my own accounts for some of them.
National chains were absent from the answers in the coffee, pizza and air conditioning categories. In plumbing, the big franchises showed up in Google’s sponsored ads below the answer, not in the answer itself. In DeLand, a plumber with 19 reviews was named in the answer while companies with thousands of reviews appeared in the ad slots.
The pattern held in every case I checked, with one honest exception. For a quick oil change, chains did lead. Even then, the assistants picked location by location based on reviews, and one assistant told the user to skip a specific location over a single bad review.
The sources the assistants leaned on were all third party. They included review sites, Reddit threads, local blogs, niche guides and local press. Claude, without being asked, called company self-descriptions “leads, not endorsements.”
The limits of my test
This was one or two cities on one day, not a study. Some assistants ran on my personal accounts, and the answers change from run to run. A few award claims that ChatGPT made were unverified, so I do not repeat them as fact. Different assistants also chose different first picks for the same coffee question, which tells me there is no single ranking to win.
I would not call this a law of nature. I would call it a consistent signal that is worth planning around. The businesses named were the ones that many independent sources mentioned for that exact job.
Putting the two lessons together
My pages worked because they were specific and because they sat on top of a real business with real customers. Specific pages get you retrieved. They do not, by themselves, get you recommended.
A page that restates your own claims is a lead, not an endorsement. If the only place a fact appears is your own site, an assistant has no reason to treat it as settled. If the same fact appears in your page, a customer review and a trade forum, the assistant has corroboration.
That is why I treat scaled content as the first half of the job. It makes you findable for the long questions. The second half happens off your site.
Practical takeaways for marketers
- Write for the question, not the keyword. Pick the long, specific questions your buyers ask and give each one its own page. Lead with the answer, then support it.
- Keep every page useful on its own. If you cannot say what a page adds that your other pages do not, do not publish it. Repetition dilutes trust.
- Check that engines can reach you. Confirm that the search crawlers from the major engines are allowed and that your pages are indexed. Training-only crawlers are a separate decision.
- Make facts explicit and consistent. Use the same names, specifications and claims everywhere. State numbers in plain sentences that can be quoted.
- Earn third-party mentions. Reviews, forum answers, local press, comparison guides and niche directories are what assistants cite. Ask real customers, publish genuinely useful material others will reference, and keep your listings accurate.
- Measure by question, not by traffic. Ask the assistants your customers’ questions every month. Record who is named, which sources are cited and what changed.
Where this leaves us
Scaling content was the right call for a newcomer facing entrenched competitors, and the growth in indexed pages was real. But the number that mattered was not 4,089. It was how many independent sources agreed that we knew the answer.
The assistants are not rewarding the biggest brand or the loudest ad budget. They reward the source that is best documented, by more people than just the source itself. Build the pages, then build the proof.
References and further reading
- Microsoft: Bing Webmaster Guidelines
