How do I improve my rankings in AI search results? What we do for clients
There is no page-one list in AI answers, only citations and mentions. This is the order we work in with clients, from crawler access to content and measurement.
When a new client asks how to improve their rankings in AI search results, the first thing we do is reset the question. ChatGPT, Gemini and Perplexity do not return a ranked list. They write an answer, cite a few sources and name a few brands. Improving "rankings" means getting cited more often and named more often for the questions your buyers ask. That is measurable, and it responds to specific work.
Here is the order we run it in, and why.
Check access before anything else
On most audits, the first real finding is not content. It is access. Bot protection, CDN rules or an old robots.txt line blocks one of the AI crawlers, and the site never gets a chance. The ones worth knowing are OAI-SearchBot and ChatGPT-User from OpenAI, ClaudeBot and Claude-SearchBot from Anthropic, and PerplexityBot. GPTBot is the training crawler, and whether you allow it is a separate business decision.
We do not trust robots.txt as proof. We want logs showing the bots arriving and getting a 200. In our client programs we use the crawler logs in Promptwatch for this, because they are IP-verified, so spoofed user agents get filtered out, and each page shows a citation rate next to its crawl history. A page that is crawled often and never cited has a content problem. A page that is never crawled has an access problem. Different fix, and you save weeks by knowing which one you have.
Build the prompt set like a buyer would
Next we agree a prompt set with the client. We include organic questions, brand questions and competitor comparisons, then tag them by topic and country. Resist the urge to track hundreds of prompts on day one. A tight set you review every week beats a huge one nobody reads.
We also look at query fan-outs, the narrower searches a model runs behind a single question. They tell you what your page actually has to answer. If a model searches for "pricing 2026" and "official documentation" when someone asks about your category, a page with neither will not be picked, however good the prose is.
Fix the pages the models skip
When the data shows a lost prompt, we look at the pages that win it. Patterns we see again and again: the winning page answers in its first lines, it has a table or a clear list, its numbers are dated, and it covers the exact sub-question. The losing page opens with three paragraphs of context.
Our usual edit list:
- Move the direct answer to the top of the section.
- Add a comparison table with figures the client can source.
- Refresh pricing and feature pages, which models cite for official lookups.
- Split broad pages into pages that match specific fan-outs.
- Pursue third-party mentions where the citation reports show them: review sites, Reddit, YouTube.
The citation analytics in Promptwatch shows which domains and pages the models cite for each prompt, so this list is driven by evidence, not taste. Its Action Board then turns findings into a prioritized to-do list we can assign, which keeps weekly work from drifting.
Close the loop with traffic
It is easy to celebrate more mentions and miss whether anyone clicked. Visitor analytics in Promptwatch attributes AI-referred visits and conversions, so we can show the client whether a gain in visibility showed up as sessions. For clients who want help producing the pages, the Content Agent on Business plans can plan and publish against gaps, and we keep a review step in front of anything that goes live.
What we tell clients about limits
We never promise a recommendation. Answers change from run to run and vary by country, so a single screenshot proves little. What we promise is a method: a stable prompt set, weekly measurement, and fixes tied to evidence. Over several weeks the share-of-voice trend tells you whether the work is landing.
If you are starting from zero, do these four things this week. Confirm the AI crawlers can fetch your key pages. Write down thirty or so buyer prompts. Find out who gets cited for them today. Then pick the three biggest gaps and fix those pages first. That sequence is dull, and it works better than any trick.