What to fix first for AI search
A running order based on what the evidence supports rather than what is easiest to sell. Three gates before anything else, then off-site work, then your own pages — which come later than almost every agency will tell you.
Most AI search checklists are ordered by what an agency can bill for. This one is ordered by evidence strength, which produces a noticeably different sequence — your own website content lands about two thirds of the way down.
Step 1: the three gates
These are conditions, not improvements. If one is shut, everything below it is wasted. All three are Grade A, documented by the providers themselves, and checkable in an afternoon.
- Bing indexation. ChatGPT discovers URLs through the Bing Search API, and Copilot is Bing-native. If you are not in Bing, two of the five surfaces structurally cannot cite you. Submit through Bing Webmaster Tools.
- Crawler access. Check robots.txt for OAI-SearchBot, GPTBot, ChatGPT-User, PerplexityBot, Perplexity-User, ClaudeBot and Google-Extended. They do different jobs: blocking OAI-SearchBot removes you from ChatGPT's citation eligibility, while blocking GPTBot only removes you from training.
- Server-side rendering. Vercel and MERJ observed 500 million GPTBot fetches and found zero evidence of JavaScript execution. If your content only appears after JavaScript runs, retrieval sees an empty page.
Do not pay a retainer to anyone who has not checked these three things first. They are binary, verifiable, and they are the only items in this whole field where the provider documentation says plainly what happens.
Step 2: your Google Business Profile, if you are local
Google explicitly names Google Business Profile and Merchant Center as inputs to its generative AI features. That is the clearest documented lever a local business has, and it costs nothing.
Make sure the profile exists, is verified, has the right category, and that your name, address and phone number match everywhere else you appear. Inconsistent details across directories directly degrade what assistants say about you.
Check that the profile Google returns for your name is actually you. When we ran this on our own business, the closest match Google offered was a different company with a similar name. Anyone asking an assistant about us was being pointed at somebody else.
Step 3: the sources that answer your buyers' questions
This is the step most checklists put last, and the evidence puts it here. In our own audit, 59 of the 108 citations on the 'who can I hire' question were third-party roundup lists. Five of the twelve most-named companies sat on a single directory the assistants read.
Independently, Ahrefs found unlinked mentions of a brand elsewhere on the web correlate with AI visibility roughly three times more strongly than backlinks do. Both are correlational, and they point the same way.
- Run your buyers' actual questions and record which pages get cited. That is your target list, and it is specific to your category.
- Submit to the cited sources that accept listings.
- Pitch the editorial roundups.
- Answer real questions in the communities your buyers use, under your own name and with your affiliation stated.
Step 4: now your own pages
Being retrieved is not being kept. AirOps found ChatGPT cites roughly 15% of the pages it retrieves, across 548,534 pages — the other 85% are fetched, assessed and discarded. Your page content decides which side of that you land on, which is why it matters, and why it matters after you are reachable at all.
Prioritise in this order:
- Pages answering the questions your buyers actually ask, in the words they ask them. In our category, explanatory and how-to pages took 62% of all citations and case studies took none.
- Clear factual statements with the numbers and sources visible in the HTML. Answer-first structure with statistics and quotations tested at +22% to +41% in the GEO-bench study, though on a simulated engine and a proxy metric.
- Freshness. 76.4% of the most-cited ChatGPT pages had been updated within the previous 30 days. Correlational, cheap, low-risk.
- Say plainly what you do and where you do it, in crawlable text. A homepage that is mostly imagery and a phone number gives retrieval nothing to work with.
Step 5: the cheap technical items, done last and unpaid
Schema markup, llms.txt, and the rest of the technical checklist belong here: worth doing because they are correct and take an hour, not worth a line item.
Schema is the honest illustration. It is the most-recommended intervention in the category and the only one with a large-scale causal test behind it, and that test found −4.6% on Google AI Overviews and statistical noise elsewhere. Google's own guidance says structured data is not required for generative AI search.
What this order costs you if you get it wrong
The common failure is starting at step 4 or 5 because they are the visible, billable work, while a gate at step 1 is shut. A business can publish twenty excellent pages, mark them all up perfectly, and remain structurally invisible to ChatGPT because it is not in Bing's index.
The second failure is stopping at step 4. Your own pages cannot get you named on a question the assistant answers by reading somebody else's list.
Common questions
- What should I fix first to show up in AI search?
- The three gates, in this order: Bing indexation, AI crawler access in robots.txt, and server-side rendering. They are conditions rather than improvements — if any one is shut, nothing else you do can work, and all three are documented by the providers rather than inferred.
- Does adding more pages to my website improve AI visibility?
- Less than most advice implies, and later in the order than you would expect. In our own measurement the majority of what decided whether a company was named sat on third-party pages. Publishing more of your own pages is the slower half of the job.
- How do I know if ChatGPT can even see my website?
- Check two things. First, whether Bing has your site indexed, because ChatGPT discovers URLs through Bing. Second, whether your robots.txt allows OAI-SearchBot specifically — blocking it removes you from citation eligibility even if you allow every other crawler.
Sources
Every claim above should be checkable. Where a study has limits, they are stated rather than left out.
- Google Search Central — generative AI guidance, Business Profile as an input, structured data not requiredGrade A. Provider documentation.
- Vercel and MERJ — 500 million GPTBot fetches, no JavaScript execution observedGrade A/B.
- AirOps — 548,534 pages across 15,000 prompts, retrieval versus citationGrade B. Large sample.
- Ahrefs — schema difference-in-differences causal study; and 75,000-brand correlation studyGrade A for the causal null result; Grade B for the correlations.
- GEO-bench (Princeton, Georgia Tech, IIT Delhi, Allen AI), KDD 2024Grade B. Simulated engine, proxy metric.
- Sentinay — original measurement, 29 answers and 287 citation URLs, 17 August 2026Grade B. Our own single run.
More on getting cited in ai answers
- Why the five AI assistants disagree about who to recommend
We measured every source five assistants used to answer the same questions. Not one domain was cited by all five, and 89% were cited by only one. Our own original data on why a single AI visibility number is misleading.
- The source ecosystem: how AI actually decides who to name
When we asked five assistants who to hire in our own category, 59 of 108 citations were other people's roundup lists. Five of the twelve most-named companies sat on one directory. Being named is mostly about other people's pages.
- What actually works in AI visibility, graded by evidence
Twenty interventions sold as AI visibility work, each rated by how strong the evidence behind it actually is. Three are gates you must pass. One has a rigorous causal study behind it, and that study found nothing.
- How AI assistants pick which businesses to name
ChatGPT, Perplexity, Google's AI answers and Copilot use four different pipelines to decide who gets recommended. An intervention that works on one can do nothing on another, which is why they have to be measured separately.
- Does llms.txt do anything?
Over 844,000 sites have adopted llms.txt. No major AI provider has committed to reading it in production retrieval, and the largest correlation study found no relationship with citations. Here is what it is actually for.
Want to know where you stand?
We ask five AI assistants for a company like yours and send you a free report showing how often you came up and who came up instead.
Get my free report