4 minute read

Key takeaways

  • No: no published evidence establishes that an LCP above 2 seconds is “disqualifying for ChatGPT”.
  • The claim circulates on the market; OpenAI’s documentation governing its crawlers’ access, read on 19 August 2026, mentions neither speed nor any loading threshold.
  • On our measured terrain, where speed and citations are read side by side, the most-cited site carries the worst speed score of the panel, and the fastest site is barely cited.
  • Speed remains genuinely worth working on, for your visitors and for machine readability. What lacks evidence is the precise threshold presented as a guillotine.

In this article

A business owner showed us this rule, read on a provider’s website: beyond 2 seconds of LCP, it would be “disqualifying for ChatGPT”. The rule is precise, it is frightening, and it steers budgets. So we did what our charter requires: we looked for the evidence. Here is what we found, sources and dates attached.

Where the claim comes from

It appears in an article published on 20 March 2026 by L’Agence WAM, accessed 19 August 2026: “LCP (temps de chargement des pages) sous 1 seconde (au-delà de 2 secondes, c’est rédhibitoire pour ChatGPT)”, that is, beyond 2 seconds it would be disqualifying for ChatGPT. LCP, Largest Contentful Paint, measures how long the largest visible element of a page takes to display; it is one of Google’s three Core Web Vitals signals.

Let us be precise about what we are not saying: the same article contains interesting field observations, and it itself acknowledges uncertainty on a neighbouring topic, structured data, “even though we are not certain ChatGPT takes it into account” (our translation). Our question is about one thing only: the 2-second threshold presented as eliminatory. A rule this precise deserves evidence this precise.

What OpenAI’s documentation says, and does not say

OpenAI’s page explaining how a website becomes readable by its robots, developers.openai.com/api/docs/bots, read on 19 August 2026, states one condition: allow the robots, including OAI-SearchBot for search, in the robots.txt file. It mentions neither LCP, nor speed, nor performance, nor any loading threshold. An absence in the documentation does not prove a factor does not exist; it proves no vendor has published it. Between “undocumented” and “disqualifying” lies all the distance our trade must respect.

What a measured terrain shows

Our public diagnostic of 7 and 8 August 2026 measures, on the same dates and with the same tools, the speed of five websites of one market and their presence in 270 AI answers. The four compared schools are anonymised, as in the report. Here are the two columns side by side, mobile PageSpeed of the homepage and answers linking the domain:

SiteMobile PageSpeed, of 100AI answers linking the site
Audited site (Real French)966 of 270
School A3934 of 270
School B9228 of 270
School C560 of 270
School D5210 of 270

On this terrain, the most-cited site of the panel carries the lowest speed score, 39 of 100. The fastest site, 96 of 100, is linked in only 6 answers of 270. If the 2-second threshold were eliminatory, this table would be impossible.

The limits of this finding, stated plainly: one market, five sites, two dates. It does not prove speed never matters; a page so slow that a robot gives up reading it will be read by nobody. It proves that a precise 2-second threshold, presented as a guillotine, survives neither the vendors’ documentation nor this measured terrain.

So, should you work on speed?

Yes, and we do it for our own websites: speed serves your visitors, it is a documented signal in classic Google search, and display stability is part of agent-readability checks. That is why we publish the measured speed of the market’s websites, dated captures included, ours as well. Do speed for the right reasons; do not do it out of fear of a guillotine nobody has measured.

Our rule, and the usual invitation

Our charter forbids us to assert without a date and a source, and obliges us to correct publicly. The same door is open here: if a published measurement establishes an eliminatory speed threshold for ChatGPT, write to us, we will read it and this page will be corrected, with the date.

Addendum of 19 August 2026, evening: the door served the very same day. A careful reading pointed us to an independent analysis of 107,000 pages (Dan Taylor, Search Engine Land, 13 January 2026, accessed 19 August 2026): it measures a weak, negative link between LCP and visibility in answer engines, a correlation of -0.12 to -0.18 depending on the definition used, visible at scale but weak in practice, and with no threshold at all. That is exactly the nuance this page defends: speed matters a little, among many other things; the 2-second guillotine remains without evidence.

Frequently asked questions

Does LCP matter for being cited by ChatGPT?

No OpenAI documentation says so, and our measured terrain shows the slowest site of the panel as the most cited; an independent analysis of 107,000 pages measures only a weak link, with no threshold at all. What OpenAI does document: let its robots read the site. The rest is hypothesis, and should present itself as such.

Is a slow website still a problem?

Yes, for your visitors first, and for classic Google search where speed signals are documented. That is reason enough to build well, without needing a scarecrow.

Why cite the source of the claim?

Because that is the rule we apply to everyone, ourselves included: a claim is quoted word for word, with its source and its date, so anyone can read it in full and make up their own mind.

Sources