What changes should I make to improve my AI search rankings?
Not a philosophy — a list. Fifteen specific changes ordered by what they actually move, who has to do them, and how long before you can tell.
This is a list rather than a philosophy. Fifteen specific changes, ordered by what they actually move, grouped by who has to do them, with an honest note on how long before you can tell whether they worked.
Tier 1: binary fixes
These are either right or wrong, and when they are wrong they are usually the entire problem. Do them first regardless of what else is scheduled.
1. Serve the answer in raw HTML. AI crawlers do not execute JavaScript. An analysis by Vercel and MERJ covering more than 500 million GPTBot fetches recorded zero JavaScript execution — it downloaded JavaScript files roughly 11.5% of the time and never ran them, with the same behaviour from ClaudeBot and PerplexityBot. Effect: invisible to visible. Timeline: days to weeks after recrawl.
2. Unblock AI crawlers at the CDN. Check robots.txt, then the bot-management preset where the block usually lives. Verify from outside your network. Effect: same magnitude as above. Timeline: days to weeks.
3. Fix server response times. AI crawlers use tight timeouts, often one to five seconds. Slow responses can mean no crawl at all, silently. Effect: moderate to high. Timeline: immediate on next crawl.
4. Confirm the pages are indexed. Retrieval leans on organic signals; an unindexed page is not in any candidate set. Effect: prerequisite. Timeline: weeks.
Tier 2: structural edits
Editorial changes to pages that already exist. These produce the most reliable improvement per hour once tier one is clean.
5. Move the answer to the first two sentences of every section. Models lift passages, not pages, and a buried answer loses to someone else’s opening line. Effect: high. Timeline: two to six weeks.
6. Make headings question-led. Phrase them the way a person asks an assistant, because that is what retrieval matches against. Effect: moderate. Timeline: two to six weeks.
7. Split pages that answer six questions. One complete answer per page beats six partial ones. Effect: high. Timeline: four to eight weeks.
8. Make passages self-contained. Name subjects instead of using pronouns that refer back. Test by copying three sentences out of context. Effect: moderate to high. Timeline: two to six weeks.
9. Add one plain entity sentence — category, audience, alternatives — to your homepage and mirror it in Organization schema. Effect: moderate. Timeline: four to eight weeks.
10. Correct structured data so it mirrors the visible text exactly. Mismatch is worse than absence. Effect: supporting. Timeline: weeks.
Find out which tier is your problem
A free scan checks retrievability and shows what assistants actually say about your category — so you start at the right tier.
Run my free scanTier 3: content investment
11. Add specific, dated, attributable claims to your key pages. The Princeton and IIT Delhi GEO study found Statistics Addition, Quotation Addition and Cite Sources were the three highest-performing methods tested, delivering roughly 30–40 percent relative visibility improvements. Effect: high. Timeline: three to eight weeks.
12. Publish one piece of original data. A survey, a benchmark, an analysis of your own operations. Being the only source for a claim means any answer using it must cite you. Effect: highest and most durable. Timeline: one to two quarters to compound.
13. Write the pages that close your prompt gaps. Take the questions where competitors are cited and you are not, and write the page that answers each one better. Effect: high, targeted. Timeline: four to twelve weeks.
Tier 4: corroboration
14. Get accurately described on the domains assistants already cite. Derive the target list from your own scan results rather than guessing — the cited URLs are a ranked map of what influences your category. Effect: highest ceiling. Timeline: one to two quarters.
15. Make third-party descriptions consistent. Same category noun, same product names, same positioning, everywhere. Ten sources agreeing beats fifty disagreeing, because inconsistency gives a model nothing stable to state. Effect: high. Timeline: ongoing.
Changes that produce nothing
Worth naming explicitly, because each of these consumes real capacity.
- Repeating your brand name. Models reason about entities, not string frequency.
- Publishing volume. Fifty thin pages give a model fifty things not worth quoting.
- Adding schema to unreadable pages. Markup on a page nothing can fetch changes nothing.
- Submitting your site. There is no form.
- Buying placement in organic answers. It does not exist. Advertising inside AI surfaces is labelled and separate from cited sources.
- Keyword density work. Engines match meaning, not counts.
- Changing dates without changing content. Transparent and ineffective.
Who does what
| Tier | Owner | Typical blocker |
|---|---|---|
| 1 — binary fixes | Engineering | Backlog priority |
| 2 — structural edits | Content or SEO | Editorial capacity |
| 3 — content investment | Content plus subject expert | Access to original data |
| 4 — corroboration | PR or communications | No route between marketing and PR |
The most common structural failure is marketing owning the outcome without the engineering access needed to clear tier one. If a rendering change takes nine months to schedule, everything below it is theatre — and that is a resourcing conversation rather than an SEO one.
Keeping a change log
Because effects take four to eight weeks to appear and answers vary between runs, attribution is impossible without a record of what changed and when.
Keep five columns: date, change made, tier, pages affected, and the prompt-set metric at the time. It takes a minute per change and it converts an unanswerable question — “did the restructure work?” — into a lookup. It also prevents the most expensive failure mode in this work, which is making several changes at once and never learning which one mattered before applying the same approach to another four hundred pages.
How much of this applies to you
Fifteen changes is more than most teams can schedule at once, and not all of them are equally relevant to every site.
If your site is server-rendered and your crawlers are open, skip tier one entirely after verifying it and start at tier two. A meaningful minority of sites are already fine here, and they should not spend a sprint confirming it repeatedly.
If your organic search is weak, tiers two and three will underperform because retrieval depends on ranking. Fix the foundation first; AI-specific work on an unranked site is premature.
If you are already cited regularly, your remaining upside is almost entirely in tier four and in change 12 — original data. The mechanical and structural work has a ceiling and you have reached it.
Establishing which of these describes you takes one baseline scan and one JavaScript-disabled check, which is why both sit in week one of the plan below rather than anywhere later.
Verifying each change landed
Every change on this list can be verified, and skipping verification is how teams end up believing they fixed something that is still broken.
Tier one is verifiable immediately. Fetch the live URL with JavaScript disabled and confirm the content is present. Request the page from outside your network with an AI crawler user agent and confirm a 200. Both are objective and take a minute. Caching layers routinely serve the old version after a deploy, so check the live URL rather than trusting the ticket.
Tier two is verifiable structurally. Copy three sentences from the middle of the edited page and read them cold — if they answer something on their own, the edit worked. This does not require waiting for a recrawl.
Tiers three and four are only verifiable by scanning. Which is why the fixed prompt set and the stored history matter: without them you cannot distinguish a change that worked from ordinary variance between runs.
The most common false conclusion
“We made the changes and nothing happened” is most often reported at week two or three, which is before any engine except Perplexity could plausibly reflect them. Check Perplexity first for early confirmation, then wait the full four to eight weeks before judging ChatGPT.
A ninety-day order
- Week 1. Baseline twenty prompts across three engines, twice each. Run the JavaScript-disabled check and the crawler check.
- Weeks 2–3. Fix everything in tier one. Nothing else until this is verified.
- Weeks 4–8. Tier two on your ten most important pages, prioritised by the prompts where competitors are cited and you are not.
- Weeks 6–12. Tier three, starting with adding specificity to existing pages and one original data piece commissioned in parallel.
- Week 4 onward, continuously. Tier four. Start early because it takes longest, and derive the target list from your baseline scan.
- Weekly throughout. Re-scan the fixed prompt set. Read trends, never single runs.
Tier four running in parallel from week four rather than sequentially at week twelve is the one scheduling decision that most affects the twelve-month result, because corroboration is the slowest thing to compound and the easiest to postpone.
Sources
- Vercel and MERJ — analysis of 500M+ GPTBot fetches finding zero JavaScript execution; GPTBot downloaded JS files ~11.5% of the time without running them; same behaviour for ClaudeBot and PerplexityBot.
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande — “GEO: Generative Engine Optimization”, ACM SIGKDD 2024 (arXiv:2311.09735); ~30–40% relative visibility improvement across ~10,000 queries.
- SERP feature tracking, 2026 — AI Overview presence averaging ~32.5% of monitored queries within a 31.02%–34.40% band.
Frequently asked questions
What changes should I make to improve my AI search rankings?
In order of leverage: make sure your answers exist in raw HTML and AI crawlers are unblocked, which are binary and often the entire problem; restructure pages to answer one question in the first two sentences; add specific, dated, attributable claims and original data; then build consistent third-party description on the domains assistants already cite in your category.
What is the single highest-impact change?
Making sure the answer exists in raw HTML. AI crawlers do not execute JavaScript — Vercel and MERJ recorded zero JavaScript execution across more than 500 million GPTBot fetches. Where this is the problem, fixing it moves you from invisible to visible, which is a larger change than any editorial improvement can produce.
How long before changes show up in AI answers?
Perplexity often reflects a change within days. Gemini and AI Overviews follow Google’s recrawl cadence, typically a few weeks. ChatGPT is slowest because it also depends on third-party sources updating. Plan on four to eight weeks for a readable trend, and never judge on a single answer.
Do I need to change my whole site?
No. Retrievability fixes are usually template-level and apply everywhere at once. Editorial changes should be targeted at the specific pages that should be answering the prompts where competitors are cited and you are not — typically ten to twenty pages rather than the whole site.
Which changes do not help?
Repeating your brand name, publishing high volumes of thin pages, adding schema to pages nothing can read, submitting your site anywhere, and buying placement in organic answers, which does not exist. Also stuffing keywords, since models match meaning rather than counting strings.
Should engineering or marketing own these changes?
Both, with a single owner coordinating. Tier one is engineering — rendering and crawler configuration. Tiers two and three are editorial. Tier four is closer to PR. The common failure is marketing owning the goal without the engineering access needed to clear the first tier.