Guide
Generative engine optimization for software: get cited by AI assistants
Nobody controls what an assistant says about a product. What you can control is whether the assistant can reach your pages, read the facts on them, and find the same facts confirmed elsewhere. This is a practical checklist for software makers, with one dated observation from our own sites.
By the Toolfound team. Last reviewed
What generative engine optimization means here
Generative engine optimization, also called AI search optimization, is the work of making a product easy for an AI assistant to find, read and describe correctly when someone asks about it or about its category. It overlaps with ordinary search optimization, because the same crawlable, well-structured, accurate pages serve both. It differs in the output: an assistant composes an answer and may name a few sources, instead of ranking ten links.
No tactic guarantees a mention. Anyone who sells guaranteed placement in assistant answers is promising something no one controls. The checklist below is about removing reasons a good product gets passed over.
What we observed: assistants read far more than they send
Between 7 September and 4 October 2026, a 28 day window, we compared two numbers across the 15 sites we operate. They measure different things and should not be added together.
| Measure | Source | Count |
|---|---|---|
| Page fetches by ChatGPT's user-initiated browsing, classified as verified | Cloudflare request data | about 5,100 |
| Of those, on toolfound.com | Cloudflare request data | 1,059 |
| Sessions that arrived with a ChatGPT referrer | Our web analytics | about 130 |
In September 2026 we observed roughly forty times more fetches than referred sessions. The comparison has limits. Our analytics only sees visits that run its script and keep a referrer, so it undercounts. Counting visits that carry a ChatGPT campaign tag as well as a referrer raised the portfolio figure to 521 over 30 days, which is still about a tenth of the fetch count. Many fetches are of a site's home page when someone asks about a product by name. The numbers describe our sites in that window, not a benchmark for yours.
The practical lesson is modest and useful: an assistant reads your pages much more often than it sends a visitor, so the text it reads has to carry the facts. A click-through is not the only thing to measure.
The checklist
| Check | What to do | How to verify |
|---|---|---|
| 1. Plain statement | Say on the home page what the product is, who it is for and what it costs, in one or two sentences. | Read only the first screen of text. Would a stranger repeat it correctly? |
| 2. Let fetchers in | Do not block assistant crawlers or browsing fetchers in robots.txt, a firewall or a bot-protection rule. | Request your pages with curl using a generic user agent and read robots.txt. |
| 3. Facts in the HTML | Put price, features and limits in server-rendered text, not only in scripts or images. | Run curl on the page and search the output for the price. |
| 4. One page per question | Give pricing, how it works, integrations and comparisons their own pages. | List the questions customers ask and match each to a URL. |
| 5. Matching structured data | Add JSON-LD such as SoftwareApplication with offers, and FAQPage only for real questions on the page. | Parse the JSON and compare each value with the visible text. |
| 6. Consistent footprint | Use the same name, description and URL on your listings, directories and docs, and verify ownership where you can. | Search for your name and read what third-party pages say. |
| 7. A machine surface | Offer an API or MCP server with described tools, and optionally an llms.txt. | Run the MCP listing preflight and the llms.txt validator. |
| 8. Measure reads and visits apart | Count fetches from edge or server logs by verified user agent, and visits from analytics with tagged links. | Compare both counts for the same dates. |
| 9. Defensible claims | Date your statistics and keep claims you can source. Assistants repeat what the page says. | Ask who could check each number on the page. |
The checks that go wrong most often
Facts that exist only in JavaScript
Many fetchers retrieve HTML and do not run scripts. A single-page app that renders its pricing in the browser can look empty to them. View the page source, or request it with curl, and check that the facts are present before judging anything else.
Bot protection that blocks the wrong visitors
A challenge page returned with a 200 status looks like a normal page to a fetcher. If you use a firewall rule against automated traffic, test that a polite, identified fetcher still gets your real content.
Structured data that disagrees with the page
Markup is a second copy of your facts. If it lists a price the page does not, it is a source of contradiction. Generate it from the same data as the visible text.
A footprint nobody can corroborate
When several independent pages describe a product the same way, an assistant has more to go on. Ownership-verified listings help because the facts come from the owner, which is the reason Toolfound verifies ownership before a listing goes live. We cannot promise any assistant will cite a listing.
What to avoid
- Hidden text, keyword stuffing or invisible instructions aimed at models.
- Fake reviews, invented statistics or numbers you cannot source.
- Hundreds of near-duplicate pages with a swapped noun. They dilute the pages that matter.
- Promises or purchases of guaranteed placement in assistant answers.
How to tell whether it is working
- Write down ten questions a buyer would ask about your category and your product by name.
- Ask them in a fresh session of the assistants your customers use, and record the date and the answers. Answers vary from run to run, so repeat a few times.
- Check whether the product is named, whether the facts are right, and which sources are cited.
- Fix wrong facts at the source page, not by arguing with the assistant, and repeat the questions next month.
- Keep the fetch count and the visit count side by side, as in the table above.
The MCP listing preflight and the llms.txt generator cover item 7, and the redirect checker shows whether the links you list still land where you expect.
Background reading: llms.txt explained covers what that file can and cannot do, and the guide to MCP server directories compares where an MCP server can be listed.
Common questions
How is generative engine optimization different from SEO?
The groundwork is shared: crawlable pages, accurate facts and honest structure. The difference is the output. A search engine ranks links, while an assistant writes an answer and may cite a few sources, so the clarity and accuracy of the text matter more than position.
Do I need an llms.txt to be cited?
No. It is an optional convenience and we know of no major provider that has committed to using it. The llms.txt guide on this site covers what is and is not known.
How do I know whether assistants mention my product?
Ask your own category questions in fresh sessions and record dated answers, repeating a few times because answers vary. Watch your server or edge logs for verified assistant fetches, and your analytics for referred visits.
Why are fetches so much higher than referred visits?
Assistants fetch pages to answer a question and often show the answer without a click. Analytics also misses visits that block scripts or lose the referrer. The ratio we saw describes our sites over one window, not a rule.
Can I pay to be recommended by an assistant?
Be wary of anyone who says so. No vendor controls what an assistant writes. Spend the effort on pages, structure and corroboration instead.
How long until changes show up?
It depends on the assistant and how it retrieves information, so there is no reliable timeline. Re-ask your ten questions on a regular schedule and keep a dated log of what changed.