If you built something small and put it online this year, you have probably noticed that the old advice stopped working. You wrote the page, you picked the keywords, and the traffic did not come. Nothing is broken. The place people ask their questions has moved.
Google's AI Overviews now reach billions of people a month, and its AI Mode passed a billion users inside its first year. More than 60% of searches end without anyone clicking anything. The person who needed your tool got their answer, and the answer was assembled by a model that read about you rather than sending anyone to you.
Three jobs, not one
It helps to stop thinking of this as one task. Classic search optimisation makes your page eligible to be found at all. Answer engine optimisation makes the specific fact someone asked for easy to lift out of the page. And what people are calling generative engine optimisation is about whether your name is something a model is willing to say out loud, which comes down to whether you look like a real, consistent, checkable thing.
You need all three, but they fail in different ways, and the third one is where small makers lose without noticing.
The failure nobody sees
Here is the one worth checking today, because it is invisible from your own homepage. Everything a machine reads about you lives in places you never look: your structured data, your llms.txt if you have one, the description buried in your page source. Those layers drift away from what your site actually says, and nothing warns you.
We had exactly this problem on this site. Our homepage said clearly that the marketplace was pre-launch and nothing could be bought yet. Our structured data said the opposite: every product carried a machine-readable claim that it was in stock and purchasable today. Both were written by us. Neither was lying on purpose. They simply lived in different files and drifted.
A person reads your homepage. A machine reads four other things. If those disagree, the machine wins, because it is the one doing the talking.
What actually helps
- Answer the question in the first sentence, then explain. Models lift the sentence that stands alone.
- Put your facts in structures: a real FAQ, a comparison table, a short list. Prose buries the answer inside a paragraph.
- Make sure every marked-up answer appears in the visible text, word for word. Markup that describes a page you did not write is the fastest way to get ignored.
- Add an llms.txt: a plain text page saying what you are, in sentences you would be happy to see quoted back.
- Say the same thing about yourself everywhere. A model that finds three different descriptions of you trusts none of them.
- Check the machine-readable half every time you change the human half. They drift silently.
Credibility beats volume now
The old game rewarded whoever accumulated the most links. Language models weigh that far less. What they respond to is whether your content reads as credible, current, and specific enough to quote. A page that says something concrete and checkable is worth more than five pages of competent filler, which is a genuinely good change for anyone building alone.
It also means the small maker is not automatically at a disadvantage. You cannot outspend a company on links. You can absolutely be clearer, more specific and more honest than one, and that is now the thing being measured.
None of this is a trick, and anyone promising you a ranking is guessing. What you can control is whether the machine reading you finds something true, specific and easy to repeat. Start with the layers you never look at: ours are our llms.txt and the structured data on every listing page, both public and both worth stealing the shape of.