Why latency matters for AI search

Site speed is usually discussed as a Core Web Vitals question, and for Google rankings it matters at the margins. For AI answers I think it matters a great deal more. Assistants answer from what they already know whenever they can, because every search costs time and money. When they do go to the web, they are assembling an answer while someone waits, and a page that is slow to respond can simply be left out.

This post sets out the evidence for both halves, and what to fix. The wider picture of how crawlers fetch and render pages is in the technical SEO guide.

Every major assistant decides, prompt by prompt, whether to search at all. Anthropic's documentation says Claude searches when a request depends on information that is current, changing or outside its training data, and answers directly when the request draws on stable knowledge.¹ OpenAI says its models can choose to search the web or not based on the prompt,² and that ChatGPT may search automatically when a question would benefit from current information.³ Google's Gemini API describes the model analysing the prompt and deciding whether a Google Search can improve the answer.⁴

Google was open about the reason when it launched grounding for developers in 2024. Not every query needs grounding, and grounding adds cost and latency, so a dynamic retrieval setting gave each prompt a score between 0 and 1 and only searched above a threshold, 0.3 by default.⁵ That setting belonged to older Gemini models, but the logic has not changed: search only when it is worth the cost.

Two studies put numbers on how often that happens in ChatGPT. Semrush analysed more than a billion lines of US clickstream data from October 2024 to February 2026 and found ChatGPT using search on 34.5% of queries by February 2026, down from 46% in late 2024, with wide swings in between, from 15% to 66.3%.⁶ Nectiv ran more than 8,500 commercial prompts across nine sectors and found a search in 31% of them, rising to 41% for commerce and 59% for local.⁷ Local and shopping questions trigger searches more often because the answer depends on current details.

For the majority of prompts, then, what the model already knows about you is the answer. That is shaped by what was written about you before the model was trained, which is one reason mentions on sites people read matter, and something to weigh when you decide whether to block training crawlers.

The price lists show why systems avoid searching when they can. Anthropic charges $10 per 1,000 searches on top of token costs, counts search results as input tokens, and notes that simple factual queries typically use one to three searches while comparisons can use ten or more.¹ Its web fetch tool has no extra charge beyond tokens, and an average 10 kB page comes to about 2,500 tokens.⁸ OpenAI charges $10 per 1,000 web search calls plus the search content tokens at model rates.⁹

Time is the other cost. OpenAI notes that agentic searches take longer than quick lookups.² Anthropic's fetch tool caches pages, and its documentation discourages turning the cache off because bypassing it increases latency.⁸ A product built to answer quickly and cheaply searches only when it has to, reads as few pages as it can, and does not hold the answer back for a slow one.

Search engine crawlers have always worked to a time limit. Google's crawlers wait for content for a limited time and then pass on whatever they received, with a timeout that depends on the user agent.¹⁰ Network timeouts are treated like server errors, and indexed URLs that stay unreachable are removed from Google's index within days.¹¹

No AI provider publishes a timeout for its fetchers, so the evidence comes from tests, and the tests disagree on the number. EdgeComet observed ChatGPT-User consistently timing out beyond five seconds, logged as an HTTP 499 (client closed request), while GPTBot waited up to nine seconds. It gives no sample size, and it sells caching and rendering, so treat that as one observation.¹² RESONEO saw a ChatGPT-User request carrying a header that suggested a 15-second timeout.¹³ Five seconds or fifteen, both are measured in seconds, and both are far less patience than a person gives a page they chose to open.

Search APIs built for AI work to tighter limits still. Perplexity describes its search API as built for completeness and relevance under a tight latency budget, with prefilters that remove clearly non-responsive or stale content from the candidates, and a median latency of 358 milliseconds.¹⁴ At that scale a page is either already in the index and fast to serve, or it is out of the running.

The citation data points the same way, with the usual caveat that it shows correlation. Mike King reported Profound data on a random sample of 700,000 pages in April 2026: pages that failed to load more than 75% of the time saw roughly 18 times fewer citations from OpenAI than more stable pages.¹⁵ SE Ranking modelled 216,524 pages across 129,000 domains and found pages with a First Contentful Paint under 0.4 seconds averaged 6.7 ChatGPT citations, against 2.1 for pages slower than 1.13 seconds.¹⁶ The same study found pages with the best INP scores averaged fewer citations,¹⁶ which suggests that what counts is getting the HTML delivered quickly and reliably.

That HTML also has to contain the content. Anthropic's fetch tool does not support pages rendered with JavaScript,⁸ and Vercel's study with MERJ found that none of the major AI crawlers rendered JavaScript.¹⁷ A fast response that holds an empty shell is no better than a slow one. The method for checking that is in what crawlers see on a JavaScript site.

Not every AI answer involves a live fetch. Many come from an index built earlier, as with Perplexity's own index above. There the speed that matters is the speed at crawl time. Google raises its crawl capacity limit when a site responds consistently and its response times, including latency and time to first byte, stay stable or improve, and lowers it when the site slows down.¹⁸ Bing allocates crawl capacity based on site health, efficiency, signal quality and crawl value.¹⁹ A slow site gets crawled less, so its pages are older in the index the answers draw on.

Live fetching is growing fast as well. Cloudflare's 2025 Year in Review found crawling for user actions up more than 21 times between January and early December 2025, closely tracking OpenAI's ChatGPT-User.²⁰ Each of those requests is an assistant fetching a page because someone asked a question that needed it.

The metric to watch is time to first byte. It covers redirects, DNS lookup, connection and TLS negotiation and the wait for the server, and web.dev counts 0.8 seconds or less as good and more than 1.8 seconds as poor.²¹ Most of the web misses that mark: the 2025 Web Almanac found good TTFB on 55% of desktop sites and 44% of mobile sites.²²

None of this is exotic. It is the performance work most sites have been putting off, aimed at the server response and the HTML. For a business that wants to be named by assistants, I would now put it ahead of most content work.