Cited Layer › Measurements › Delivery cut-off
Measurement
The delivery cut-off is not quoted. It is calculated with.
Measured 21 September 2026 · ChatGPT, two accounts independently · Abu Dhabi consumer electronics
An answer engine read a retailer's published delivery cut-off, compared it against the current hour in the buyer's city, and demoted the category leader because its window had already closed.
Not quoted the promise. Computed with it. The store holding the best delivery offer in its category was pushed down the recommendation because the question was asked at the wrong time of day.
The evidence
What the engine actually wrote
The question, typed exactly as a buyer would: where to buy electronics in Abu Dhabi with same day delivery
"For the fastest delivery, I'd check Jumbo first because its published policy specifically says 2-hour delivery for eligible orders before 5 PM. Since it's currently around 8 PM in Abu Dhabi, that cutoff has passed for today, so Sharaf DG's checkout or local on-demand delivery/pickup may be more practical tonight."
Three separate operations are visible in that paragraph. The engine retrieved a cut-off from a retailer's own delivery page. It resolved the current local time in the city named in the question. Then it re-ordered its own recommendation on the result — naming the leader first, explaining why the leader no longer applied, and handing the buyer to someone else.
A second account, run in parallel in an identical mode, produced the same reasoning independently. This is not a one-off phrasing.
Consequences
Three things follow, and they are cheap to act on
A late cut-off wins the evening
Whoever closes at 17:00 is absent from every answer given after 17:00, no matter how good the promise is. On one store I measure, roughly a third of all traffic arrives between 22:00 and 04:00 — entirely outside a midday cut-off.
A 24-hour window ends the contest
There is nothing to compute against and nothing to fail. The competitor with the faster promise still loses the eight o'clock question if their window is shut.
Publish the timezone and the city
"Order before 1pm" is not computable. Same day when ordered before 13:00 GST is. Without the zone the engine either infers it or hedges, and a hedged answer names someone else.
There is a fourth, and it is the one most sites miss entirely: publish what happens after the cut-off. A buyer asking at 20:00 is a buyer the engine has to answer. If your page says nothing about late orders, the engine has nothing of yours to offer and reaches for a competitor who does. One line — ordered after 13:00 → next morning — keeps you in an answer you would otherwise be absent from.
The wider pattern
Facts are executed. Adjectives are ignored.
Across twenty tracked purchase queries on a live specialty retailer, one split held without a single exception. Questions answered by a checkable fact — who delivers to this city today, what is in stock, what does it cost — are answered from the retailer's own pages. Questions answered by a judgment — which brand is best, who is trustworthy — are answered from other people's pages, and the retailer's own site barely participates.
In the delivery run above, every single store named published a specific, checkable window:
| Store | Published promise |
|---|---|
| Wee.ae | 1 hour |
| Miles Telecom | 2 hours |
| Jumbo | 2 hours, orders before 17:00 |
| noon | 60 minutes |
| Carrefour | 120 minutes |
| iSTYLE | 4 hours |
| Sharaf DG | 24-hour express, orders before 12:00 |
Not one answer contained a general claim about being fast. "Rapid delivery across the Emirates" appeared nowhere, because it answers nothing. The number, the cut-off and the city are what get reproduced — and, as the measurement above shows, what get acted on.
Method
How this was measured, and how to repeat it
Everything here is reproducible, which is the only reason to publish it.
- Mode, held constant: Brave, temporary chat, free-tier account. Two ChatGPT accounts run in parallel in separate browser profiles, plus Perplexity.
- Control question first in every session — one whose answer is known to return a vendor list. A failed control discards the session rather than reporting it.
- Queries, verbatim: best online electronics store in Abu Dhabi · where to buy electronics in Abu Dhabi with same day delivery
- Date: 21 September 2026, evening Gulf time — which is the reason the cut-off behaviour was visible at all. The same question at 10:00 returns a different order.
- Disclosure: runs were made from a Ukrainian IP address. Answer engines weight location, so a name set reproduced from inside the UAE may differ. Both questions name the city explicitly, which reduces but does not remove the effect.
The last point is the sort of thing an audit vendor is not supposed to mention. It is here because a measurement whose conditions are hidden is not a measurement.
Caution
What this does not mean
It does not mean you can be ranked
Positions do not exist in answer engines. In a parallel measurement the same question, asked in the same minute on two accounts of identical tier and mode, returned a five-name shortlist to one and no shortlist at all to the other. Ten days earlier the same question fifty minutes apart returned five names each time, of which two matched. Anyone selling a position in an AI answer is selling something with no definition.
What can be measured is the share of answers naming you, on a fixed question set, measured identically before and after.
It does not transfer to every question
The cut-off effect lives in the fact lane. It does nothing for "which brand is best" or "who is trustworthy" — those are decided elsewhere, by sources you do not own, and no amount of work on your own delivery page moves them.
It does not depend on a manifest file
Across roughly 2,100 verified crawler fetches on two sites, requests to /llms.txt from GPTBot, ChatGPT-User, OAI-SearchBot, PerplexityBot and ClaudeBot numbered zero — only bingbot and YandexBot ever fetched it. The retrieval that produced the answer above went to ordinary pages. On the site where this was measured over a two-week window, the agent that fires on a live question took the homepage on 84% of its visits.
Check your own cut-off
Two questions, ten minutes: ask an answer engine where to buy what you sell in your city with delivery today, once in the morning and once after your cut-off. If you disappear between the two, the fix is a line of text.
If you would rather have it measured properly — ten of your buyers' questions, three runs each across two engines, sixty recorded answers — that is what the visibility audit is. The category check that precedes it is free.
alex@citedlayer.com