Four numbers frame this page speed and AEO test. In SE Ranking's data, pages with a first contentful paint under 0.4 seconds averaged 6.7 ChatGPT citations against 2.1 for pages over 1.13 seconds, which is only correlation (section 1). LCP on /citedby fell from 8.65s to 2.92s, our largest change (section 2). Pages over the CLS limit fell from 8 to 0 (section 3). The number that matters most for ThriveStack citedby is the last one. After the fixes, visibility rose from 9.4% to 27.6% and retrievals of our domain went from 33 to between 82 and 94 per run, but our prompt list also grew from 32 to 45. On the same 32 prompts, visibility moved from 9.4% to 12.2%, which is still inside our normal range (section 4).
A slow page gives an AI crawler a reason to move on. We fixed our own pages first and saved a citation baseline. Now we will publish what happened, good or bad.
Gururaj Pandurangi, FounderThriveStackThriveStack citedby Research, Page Speed and AEO (Nov 2026)
What page load speed is good enough to get cited by AI?
No AI engine publishes a speed cutoff, so no number guarantees a citation. The public data points one way, though. In SE Ranking's study, pages with a first contentful paint under 0.4 seconds averaged 6.7 ChatGPT citations. Pages over 1.13 seconds averaged 2.1. Treat anything slower than Google's good limits as a risk.
- HTML response time. Google calls 0.8 seconds or less good and over 1.8 seconds poor. A crawler waits on this before it reads a single word, so a slow server is the first place a page can lose.
- The 1 to 5 second band. Some vendors say AI crawlers give up in 1 to 5 seconds. No AI company publishes a timeout, so we show this as a claim only.
- Our slow page. The GA4 research page took 2.4 to 3.5 seconds to respond in our runs. That sits in the red zone, so it is the page we check next.
The best public number we found comes from SE Ranking, which studied 129,000 domains. The authors call the patterns directional. Stronger sites tend to be faster, so speed may be a sign of quality and not the cause.
The same study, as Search Engine Journal reports it, found one odd result. Pages with an interaction delay (INP) of 0.8 to 1.0 seconds averaged 4.5 citations, against 1.6 for pages under 0.4 seconds. The authors suggest very simple pages may signal less depth. We would not read much into that, which is why INP has no line to draw below.
So what is good enough? Use Google's good limits as your floor, because they are the only thresholds with a published source. Server response and the text in your HTML reach a crawler directly, so fix those first. LCP, CLS and INP mainly help your visitors. Our own pages show the gap: /citedby still sits in the amber zone for LCP at 2.92s.
- HTML response (TTFB): 0.8 seconds or less.
- First contentful paint (FCP): 1.8 seconds or less by Google's rule. SE Ranking's best-cited group was under 0.4 seconds.
- Largest contentful paint (LCP): 2.5 seconds or less.
- Cumulative layout shift (CLS): 0.1 or less.
- Interaction to Next Paint (INP): 200 milliseconds or less.
Try ThriveStack citedby: Check which of my pages AI can read →
How did we improve largest contentful paint on our site?
We improved largest contentful paint by removing what blocked the first paint. Each page now inlines only the CSS it uses, loads far less JavaScript and skips heavy hero images. In Lighthouse on mobile, /citedby fell from 8.65s to 2.92s. Four of the five pages we sampled got faster.
- Critical CSS. Pages had to load a 393 KB stylesheet before showing anything. Each page now inlines its own rules and loads the rest without blocking.
- Less JavaScript up front. A shared helper was pulling a chart library onto every page. JavaScript preloaded on every page fell from about 400 KB to about 150 KB gzipped.
- No hero image. We rebuilt the /citedby hero in HTML and SVG, and shrank its headshots from 685 KB to 6 KB. Total page weight fell from 1,378 KB to 727 KB.
One page went the wrong way. The GA4 research page averaged 3.83s against 3.30s before, because its server took 2.4 to 3.5 seconds to respond in these runs. Front-end fixes cannot help with that, so it is our next thing to check. /citedby is also still above Google's 2.5 second threshold in Lighthouse. Our lighter throttled probe puts these same pages at about 0.6 to 1.8s, so treat any one tool as a guide. Real-user reports use a rolling 28 days and will catch up slowly.
How did we fix cumulative layout shift?
We fixed cumulative layout shift with fallback fonts sized to match our web fonts, hero glows moved with a transform, and a few line breaks that stay put as fonts load. Pages over the 0.1 limit fell from 8 to 0 on the live site. The home page dropped from 0.293 to 0.021, across the phone widths we tested.
The home page surprised us. Fonts were not the cause. Its hero tabs change every 5 seconds at different heights, which moved a decorative glow by up to 463px. A short test never saw it, and Lighthouse did.
One page got worse before it got better. On the Semrush page at 360px, two fonts arrive about 400ms apart, so a line wrapped and then wrapped back. Our first fix took CLS from 0.165 to 0.269. A line break after the H1 colon and one in the intro, plus keeping the brand name on one line, brought it to 0.029. The good threshold for CLS is 0.1.
What happened to AI citations after the speed fixes?
Not enough to credit the fixes yet. Visibility rose from 9.4% to 27.6% across the probes on November 1, 3 and 4, but our prompt list grew from 32 to 45 at the same time. On the original 32 prompts, visibility moved from 9.4% to 12.2%, which is inside our normal range.
Raw retrievals of our domain rose from 33 on October 27 to 91, 94 and 82 on the three runs after the fixes. Semrush stayed in its usual range, with 24, 20 and 14. But eight of the 13 new prompts name us in every answer, so a bigger panel lifts the total by itself. The fair read is like for like.
On the 32 prompts in both runs, visibility went from 9.4% to 12.2% (35 of 288 answers), and answers citing our domain went from 9.4% to 11.1% (32 of 288). Three prompts were already at 100% before the fixes and still are. From October 6 to 27, visibility on each engine ranged from 3.1% to 12.5%, so 12.2% is not a signal yet.
Timing matters too. The first probe after the fixes ran on November 1, the day the last fixes shipped, so engines had little time to recrawl. We will keep probing weekly and report a verdict in December. Updated November 9, 2026.
Does page speed affect SEO the same way it affects AI answers?
Page speed does not affect SEO and AI answers the same way, and the evidence for AI is correlation so far (see section 1). Stronger sites also tend to be faster, which can explain part of the gap in SE Ranking's study.
The mechanism needs care too. Vercel found that none of the major AI crawlers run JavaScript, so a crawler never paints your page or sees a layout shift. See the quote near the top of this page. Our reading is that LCP and CLS help people, and they help AI answers only indirectly, through things like rankings and links. Server response time and text in the HTML reach the crawler directly.
Where should core web vitals work sit in an AEO plan?
Core web vitals work should sit after the basics of an AEO plan, and before polish. Make sure your answers are in the server-rendered HTML first, because a crawler can only quote what it can read. Then fix page load speed, because visitors and rankings benefit either way, whatever the citation result turns out to be.
- Find what really shifts. Our home page problem appeared only in a long test. Run Lighthouse as well as a short probe.
- Test more than one phone width. The Semrush problem appeared at 360px and nowhere else.
- Measure on the live site. One of our checks hit a redirecting host and reported the fix as missing.
- Check the server too. A slow response undid our front-end gains on one page.
- Baseline before you ship. We had citation data from October 27, which is what makes a fair test possible now.
How will we measure AI search visibility after the fixes?
We will measure AI search visibility with the same weekly prompts on the same engines, for four weeks after the fixes. We will track retrievals, citations and visibility, and compare them to the competitors in those answers. If the fixes did nothing we can detect, we will publish that too.
This test has a limit we cannot remove. The speed work shipped alongside new content on the same pages, including a richer Gemini page and a new UTM guide. Pricing text changes are coming as well. So a rise could have several causes, and we will say so when we report it.
What page speed checklist should you run before you ask AI to cite you?
Run these 14 checks in order. Start with the five crawler checks, because AI bots read your server HTML and do not run JavaScript. Then check the page is fast enough for AI to cite it: clear Google's good limits for FCP, LCP, CLS and INP, and aim for FCP under 0.4 seconds. Last, save a citation baseline and probe weekly. Tick each box as you pass it.
Page Speed Checklist: 14 checks that decide if AI can read and cite your page
Pass marks are Google's good limits from web.dev. Only the FCP stretch goal comes from AI citation data, and that is correlation. Tick the boxes as you go.
Why: AI crawlers read the server HTML and do not run JavaScript.
No AI engine publishes a speed cutoff. Use Google's good limits as the floor, measured at the 75th percentile. The best-cited pages in SE Ranking's data painted in under 0.4 seconds.
Why: one week of AI answers can swing with no change on your site.
Not sure where to start? Your fail count tells you. Zero to two fails in group A means your pages are readable, so move to group B. Three or more means fix the server and the HTML first. Our own run: the GA4 research page still fails check 2 at 2.4 to 3.5 seconds, and /citedby still fails check 7 at 2.92 seconds.
Set your own citation baseline before you ship
ThriveStack citedby runs your prompts on the engines your buyers use each week, so you can see what really moved after a change.
Start free trial →Page speed and AEO: FAQ
What page speed is good enough to get cited by AI?
No AI engine publishes a cutoff. Aim for Google's good limits: HTML response in 0.8 seconds or less, FCP at 1.8 seconds or less, LCP at 2.5 seconds or less and CLS at 0.1 or less. In SE Ranking's data, the best-cited group had an FCP under 0.4 seconds.
Does page speed affect AI citations?
It may, but the evidence so far is correlation. SE Ranking reports that faster pages averaged more ChatGPT citations. In our own test, visibility on the same 32 prompts moved from 9.4% to 12.2% after the fixes. That is inside our normal weekly range, so we cannot credit speed yet. We will report a verdict in December.
Do AI crawlers care about LCP and CLS?
Not directly. Vercel found that the major AI crawlers do not run JavaScript, so they do not paint a page or see it shift. LCP and CLS mainly help human visitors, which may help AI answers through other signals.
What should I fix first for AEO, speed or content?
Check that your answers are in the server-rendered HTML first. A fast page with its answers hidden behind JavaScript still cannot be quoted. After that, speed work is worth doing for visitors and rankings.
How long does it take to see the effect?
Lab scores change on the day you ship. Google's real-user reports use a rolling 28 days. For AI citations, plan on several weekly probes, because one week can swing widely with no change on your site.
Try ThriveStack citedby: Check my page speed against my citations →
Sources
- Vercel and MERJ, “The rise of the AI crawler”. Dec 17, 2024. None of the major AI crawlers rendered JavaScript in this data; Gemini was the exception.
- SE Ranking, “How to Optimize for ChatGPT”. Read Nov 1, 2026. Correlational study reporting 6.7 vs 2.1 average citations by FCP.
- web.dev, Largest Contentful Paint. Good LCP is 2.5 seconds or less.
- web.dev, Cumulative Layout Shift. Good CLS is 0.1 or less.
- web.dev, Time to First Byte. Good TTFB is 0.8 seconds or less, poor is over 1.8 seconds.
- web.dev, First Contentful Paint. Good FCP is 1.8 seconds or less, poor is over 3.0 seconds.
- web.dev, Interaction to Next Paint. Good INP is 200 ms or less, poor is over 500 ms.
- Search Engine Journal, “New Data Reveals The Top 20 Factors Influencing ChatGPT Citations”. Read Nov 1, 2026. Reports the SE Ranking study of 129,000 domains, including Speed Index and INP findings.
- Discovered Labs, page speed and AI crawlability. Read Nov 1, 2026. Vendor guidance that claims AI crawler timeouts of 1 to 5 seconds. It cites no primary source, so we treat it as a claim.
- ThriveStack Lighthouse and mobile probes. Oct 31 to Nov 1, 2026, live site, mobile throttling.
- ThriveStack citedby probe data for thrivestack.ai. Weekly probes, Sep 15 to Oct 27, 2026.