Every engine we asked searched before it answered, and we stored the pages it read alongside what it said. That second field turns out to matter more than the first. Here are the two answers to the same prompt, “AI cofounder for Indians”, from the same model, six hours apart.
| Run | Pages read before answering | Recommended |
|---|---|---|
| 9 | perkpilot.online, shlokvenkar.me, opsiocloud.com, gobeunicorn.com, aicreatorhub.net, aimagazine.blog | Lore AI, Replit, Cursor |
| 10 | loreai.in, the brand’s own site, ajath.in, agihouse.in, indiacostartup.in, cursor.com | Lore AI, the brand, India Costartup, Cursor |
OpenAI, gpt-5.6-luna with web search forced on, asked as a buyer in India.
Same prompt, same model, same day. The only difference is the six pages the search returned, and one of them was the brand’s own site. When the engine read it, the engine recommended it. When it did not, it did not.
Two problems that look the same from the outside
A share of voice number cannot tell these apart, and they need opposite fixes.
Never read you
The engine answered from other pages and never fetched yours. Nothing on your site can change this answer, however good the copy, because the copy was never in the room. This is a discovery problem. It is fixed by the site check, if something is stopping the crawlers, and otherwise by being on the pages the engine does read: directories, comparison posts, community threads. For the brand in these runs that list is indiacostartup.in, cofounderdekho.in, Reddit and Startup India, because those are the hosts the runs recorded.
Read you, recommended someone else
The engine fetched your site and still preferred a rival. This is a messaging problem. The fix is on your pages: state, where a summary would pick it up, the things the rivals were praised for. In these runs that was free tier, mentorship, verified profiles and India pricing.
That split, from the largest grounded run on file, says where the work is. Twenty-six answers cannot be moved by a better homepage. Two can. The plan writes itself from the ratio: get read first, then get chosen.
Why this is the field to store
Most tools in this category store who was mentioned. Some store the citations an answer showed to the user. Few store the pages the engine fetched, which is a different and larger set, and it is the one that carries the causal story. It also produces the outreach list for free: rank the hosts by how often they were read on prompts where you were absent, and you have the places to be, with the prompts as evidence.
Nothing you publish can reach an answer that never read you. Everything you publish can reach the one that did.The whole thesis, in two sentences
What we changed because of it
At the time, we changed the old app’s pages to show this split. In an Oftheard campaign today the same idea runs through the weekly report:
- Every AI answer to a tracked question is kept with the sources it used, so you can see which pages the assistant read.
- The weekly summary gives the main reason you were not named, for example that the pages the answers use do not mention you.
- The free site check comes first, so nothing on your own site stops the crawlers before any page is written.
More from the blog
Why the same AI gives a different answer every time, and what to measure instead
Two runs, the same nine prompts, the same three models, six hours apart. Not one answer came back with the same list of names. That is not a bug in the engines. It is the reason a single answer is not a measurement.
What Indian founders are told when they ask an AI for a cofounder tool
Free tier, mentorship, verified profiles, India pricing. Across six hundred answers those four attributes decided most recommendations. They are not a ranking factor. They are a page brief, and they name the page that is usually missing.