AI assistants cite pages they can fetch, read and trust. So the work, in order, is: let their search crawlers in, make sure your words are in the HTML they receive, open each page with a plain sentence that answers a question, make it obvious who you are, and get described the same way on the sites they already cite. None of it guarantees a citation, and anyone who promises one is guessing. But most sites fail the first three steps, which means the bar is lower than it looks.
What follows is the order we’d work in, with numbers from our own audits of 110 SaaS homepages, 110 Indian startups and 161 recently funded startups, plus our October check of which sites AI answers cite for 20 buyer questions.
1. Find out who gets cited now
Before changing anything, ask the question you want to be the answer to. Use ChatGPT with search on, Perplexity, Claude and Google, logged out if you can, and write down who gets named and which pages get linked.
We did this for 4080 Studio in October with 20 prompts like “design studio for SaaS startups” and “what is llms.txt”. On the 18 questions that didn’t contain our name, Brave’s AI answer cited us once and Claude never did. Directories and listicles took most of the slots: DesignRush was the single most cited source, followed by Clutch, GoodFirms, Dribbble and Reddit threads. That told us more about what to do next than any checklist.
Repeat the same prompts monthly. AI answers change from week to week, so a single check is a snapshot, not a ranking.
2. Let the right crawlers in
Each assistant uses separate bots for training and for answering, and blocking the wrong one can take you out of the answers.
| Product | Bot that decides if you can appear in answers | Bot used for model training |
|---|---|---|
| ChatGPT | OAI-SearchBot (ChatGPT-User fetches pages live when a person asks) | GPTBot |
| Claude | Claude-SearchBot (Claude-User fetches live) | ClaudeBot |
| Perplexity | PerplexityBot | Not used for training, per Perplexity |
| Google AI Overviews | Googlebot, like the rest of Search | Google-Extended, which doesn’t affect Search |
You can block training and still allow search. OpenAI says sites that opt out of OAI-SearchBot won’t be shown in ChatGPT search answers, so that’s the one to leave open if you want citations.
Robots.txt is rarely the problem. Only 2 of the 107 SaaS policies we read blocked any of the big AI crawlers. The real problem is bot protection at the edge: 6 of 110 SaaS homepages and 17 of 110 Indian startup homepages refused or never answered our clearly labelled research bot, even where robots.txt would have let it in. If you use Cloudflare, Akamai or AWS WAF, check that the bots you allow can actually load the page.
3. Put your words in the HTML
Most AI crawlers read the HTML your server sends and don’t run JavaScript. If your headline and copy only appear after scripts run, they see an empty page.
This is more common than you’d think. 16% of the funded startup homepages we read had no headline at all in the served HTML, and 12 of 93 Indian startup homepages had the same problem. Server-render, or pre-render, at least the top of every page that matters. Our AI visibility checker shows you what a crawler sees.
4. Say what you do in the first sentence
Answer engines lift short, self-contained statements. The first lines of your homepage are where they look for “what is this company”. Only 21% of the SaaS homepages we audited opened with a sentence that names the company and says what it does. For Indian startups it was 3%. Most opened with a slogan.
The fix is one sentence: your name, what you are, who it’s for. “Acme is payroll software for restaurants in the UK.” Put it in the first paragraph even if the headline above it stays clever. How we write that line.
5. Write pages that answer the questions people ask
Assistants answer questions, so they cite pages built around questions. Our only non-branded citation in October came from a comparison page, studio vs agency vs freelancer. The format that works is a heading that matches the question and an answer in the first two sentences under it, followed by the detail. That is answer engine optimisation and it carries straight over to AI answers.
Good candidates: what it costs (see our website cost page for the shape), X vs Y comparisons, “is X right for Y”, and definitions of the terms in your field. Original data helps most of all, because a number nobody else has is something an assistant has to cite you for.
6. Make it obvious who you are
AI answers mix up entities with similar names. “4080 Studio” lost to the NVIDIA RTX 4080 in Claude’s search, which is not a problem you can fix with copy alone. What helps:
- Organization schema on your homepage, with your name, logo, founders, location and founding date. 74% of SaaS homepages already have it.
- sameAs links from that schema to your profiles: LinkedIn, Crunchbase, and Wikidata if you have an entry. Only 13% of SaaS homepages link to Wikidata, and 1% of Indian ones.
- A Wikidata entry if you’re notable enough for one. It’s a public, structured record that knowledge graphs read. We made ours in 2026.
- The same description everywhere. If your LinkedIn, your directory listings and your site describe you three different ways, a model has to guess which is right.
7. Get mentioned where assistants already look
For “best X” and “who should I hire” questions, AI answers lean heavily on third-party pages. In our check, DesignRush was cited on 7 of the 20 prompts across the two engines we could test. Studios that published their own “best studios” lists also got named, because the engines passed the lists through.
So the off-site work is real work:
- Claim and complete your profiles on the directories that show up in your step 1 check.
- Correct wrong descriptions at the source. Ours turned up in one startup database as an “Enterprise Services” company doing no-code development, which is not what we do.
- Get quoted: guest posts, podcasts, original research others cite, honest answers on Reddit and Quora threads that rank.
- Ask happy clients to mention you by name on their own sites.
This is the slowest step and the one with the most lasting effect. It’s also the core of generative engine optimisation.
8. Help the search indexes find changes quickly
Assistants with search pull from web indexes, not only their own crawls. Submit your sitemap to Google Search Console and Bing Webmaster Tools, and ping IndexNow when pages change so Bing and the products built on it see updates within hours. An llms.txt is fine to add, but no major AI product has confirmed it reads one, so don’t expect it to move anything on its own.
How to tell if it’s working
Count visits from chatgpt.com, perplexity.ai, claude.ai and gemini.google.com as their own channel in your analytics, and rerun your step 1 prompts each month with the same wording. Expect weeks, not days, and expect noise.
What doesn’t work
- Keyword-stuffed “AI-optimised” copy. Assistants quote clear sentences, not dense ones.
- Hidden text or instructions aimed at models. It’s a trust risk with no upside.
- Buying “guaranteed ChatGPT placement”. No one controls what a model says. The ones selling it can’t either.
If you want to see where your site stands on steps 2 to 4 in one go, run it through the AI visibility checker.
Quick recall
What do AI assistants cite? Pages they can fetch, read and trust, that answer the question in plain words.
Which ChatGPT bot matters for citations? OAI-SearchBot. You can block GPTBot (training) and still appear.
What’s the most common failure? Copy that only appears after JavaScript runs, and a homepage that opens with a slogan.
Why do directories matter? For “best X” questions, AI answers lean on third-party lists and directories.
Can anyone guarantee a citation? No.
Written by 4080 Studio.
