AI & Technology

What a 915-to-4,089 Page Experiment Taught Me About How AI Picks Sources

By Matt Goren

When I started selling feeder insects online, the market already had decades-old incumbents with strong websites. Winning the short, obvious Google searches looked unrealistic for a newcomer. So I went where those incumbents were weaker: the longer, specific questions customers ask before they buy. 

Over about a month, the store went from 915 pages on Google to 4,089. That growth is the part people ask about. The more useful part is what I learned about how AI assistants choose which sources to trust. 

Why I stopped chasing the short searches 

A two or three word search is crowded, and the businesses that rank for it have had years to earn links and reviews. A newcomer rarely beats them head to head. But customers do not only type short phrases. 

They ask comparison questions, how-to questions and “which one should I buy for my situation” questions. Those are longer, more specific and closer to a purchase. Most sites answer them poorly or not at all. 

What the pages were 

Each page answered one specific question completely. A page compared two products for one use case, or explained one care problem, or answered one buying question in plain words. None of them were padding. 

That matters, because the volume only helped where each page was genuinely useful. Thousands of thin pages would have been noise. The discipline was one real question, answered well, per page. 

What I measured, and what I did not 

I measured how many pages Google had indexed and which questions the store started to appear for. I did not run a controlled test of how often AI assistants named the store before and after. So I cannot claim that more pages caused more AI mentions. 

What I can say is narrower. The assistants I checked tended to draw on pages that answered the exact question being asked. That is consistent with how the platforms describe their own retrieval, but it is an observation, not a proof. 

How AI assistants seem to choose who to name 

The strongest pattern I found came from a separate set of tests. I asked ChatGPT, Google’s AI and others for local businesses: a plumber in a Florida town, and coffee, pizza and air conditioning repair in a larger city. These were US tests, in a handful of places, on one day, so treat them as an illustration. 

In the Florida town, a plumber with 19 reviews was named in the answer. Companies with thousands of reviews, including national franchises, appeared only in the paid ads underneath. In the larger city, national chains were missing from almost every answer for coffee, pizza and air conditioning repair. 

Different assistants named different businesses for the same question, so there was no single ranking to chase. One assistant told its user, unprompted, that a company’s description of itself was a lead and not an endorsement. The sources the assistants leaned on were mostly third party: reviews, local press, community forums and niche guides. 

What this means for scaled content 

Volume by itself is not the lever. The assistants appear to favor the source that answers the specific question and is also vouched for by others. A large site that nobody else mentions gives them less to trust than a small one that others cite. 

So the useful version of scale has two parts. First, cover the long, specific questions in your field, one page per question, each answered completely. Second, make sure independent sources can corroborate you: reviews that mention the exact job, mentions in local or trade press, and listings that match your own details. 

What I would do differently 

I would record AI answers from the start. I checked them informally, which is why I cannot give before and after numbers. If you are going to invest in content, write down which questions you want to be named for, and re-run those questions in several assistants every month. 

I would also be stricter about quality at scale. The risk with any large content effort is publishing pages that exist for the search engine and not for the reader. Search platforms have policies against low value mass produced pages, and an assistant has no reason to cite a page that does not help. 

A practical checklist 

If you want to test this in your own business, you do not need thousands of pages. Start with ten. 

  1. List ten specific questions a customer asks before buying, in their own words. 
  2. Write one complete page for each, with real details, prices or ranges, and a plain recommendation. 
  3. Ask ChatGPT, Google’s AI and one more assistant those same ten questions in a private window, and save the answers. 
  4. Look at which sources the assistants cite, and note which ones are independent of you. 
  5. Ask recent customers to mention the specific job in reviews, and get listed where your customers already look. 
  6. Re-run step 3 monthly and compare. 

The honest bottom line 

Scaling content got my store found for questions the incumbents ignored. Whether it made AI assistants name us more is something I did not measure, so I will not claim it. What the local tests did show is that assistants lean on specific answers and independent corroboration, and that size and ad spend did not decide who got named. 

Related Articles

Back to top button