One month on, four of the six AI engines changed how much they cite

A month ago we measured how many sources an AI answer carries, engine by engine. Between the engine that names the most and the engine that names the least, the gap ran from 1 to 5.6.

We have now measured the month that followed: 3,705 answers between 2 and 31 August 2026, on the same six engines.

The gap still runs from 1 to 5.7, and not one engine changed rank. Underneath that stillness, four of the six moved, and each of them by a different lever.

The two numbers, and they do not move at the same speed

AI Coverage is the share of an engine's answers that name at least one source: it says whether there is a door. AI Density is the number of distinct sources an answer carries, counted over sourced answers only: it says how wide that door is. Their product is the average number of distinct sources an answer carries, and it is the one figure that compares across engines without further thought.

Read one month against the other, the two behave very differently. Coverage held to the tenth of a point on three engines and moved sharply on two. Density moved on all six, and it is where the month's largest change sits.

The door: AI Coverage

Perplexity, Grok and ChatGPT hold to the tenth of a point. That is what a property of the engine looks like when nothing about it changes: ChatGPT opened its answer to outside sources in 74 out of every 100 answers in July, and in 74 out of every 100 in August, over 465 answers then and 696 now.

Gemini carries the movement to read here. It named at least one source in 94.4% of its answers and now does so in 84.0%: one answer in six that credited someone in July credits no one in August. Over 605 answers, that is well past a rounding effect, and it is the first time we have seen this number move by ten points on an engine that sits at the top of the column.

Claude went the other way, from just over half of its answers to a little under three fifths.

Mistral shows the largest number in the column, and it is the one to weigh most carefully: its July figure rested on 94 answers, the engine having entered the panel on 24 July. August is the first full month over which its coverage can be read, and 59.4% over 556 answers is the reference the next wave will be measured against.

The width: AI Density

Perplexity is the month's headline. It already cited on every single answer, so its door could not open any further; it widened instead, from 13.1 distinct sources per sourced answer to 17.1. Four more names per answer, on an engine that leaves the door open every time.

ChatGPT moved the other way, and it is worth putting next to its coverage line. Its door opens exactly as often as in July, but 1.3 fewer names go through it each time, a fifth of its width.

Claude gained on both counts at once, which is why it is the engine whose net figure improved the most in relative terms. Grok and Mistral moved modestly upward. Gemini lost half a source, a small change next to the ten points it lost on coverage.

The two multiplied

Three engines carry more names than they did a month ago, three carry fewer, and the order of the six is unchanged from first to last.

The two ends of the table moved together, which is why the ratio between them barely budged: Perplexity gained 31% and Claude gained 30%, so the distance from the widest to the narrowest went from 5.6 to 5.7. A reader who only looked at that ratio would have concluded that the month was quiet. It was the busiest we have measured.

What this changes for what you do

Each of these numbers points at a different move.

Where coverage is high and stable, the question is one of competition inside a known list: Perplexity, Grok and Gemini open their answers to outside sources in more than eight cases out of ten, month after month, so the room exists and it is contested. Where coverage sits lower, as on Claude and now on Mistral, a large part of the answers is written without crediting anyone, and the number to watch first is whether that share is moving, which on Claude it is.

Density tells you how crowded the room is once you are in it. An engine at 17 names leaves space for a long list; an engine at 3.7 keeps a short one, and the same visibility is worth more there. That space is not for sale: on ChatGPT, the slot that can be bought sits below the answer, on separate systems.

And both numbers moved inside a single month, on four engines out of six. That is the argument for reading them on a live panel rather than once: a figure measured in July was already worth 20% more or less in August.

Method

The measurement covers the measurement panels running on the Epovest platform between 2 and 31 August 2026, a total of 3,705 answers across six engines. The window opens the day after the one measured by the previous wave, so the two leave neither a gap nor an overlap.

The comparison figures are those published by that wave, measured between 6 July and 1 August 2026 on 1,949 answers. Two engines are measured over a full month for the first time here: Mistral and Grok entered the panel on 24 July and carried 94 answers each in the previous window, against 556 each in this one.

A source is the registrable domain of a page the engine cites, resolved against a versioned public suffix list, so that three pages of the same site count as one source. Only pages the engine cites in its answer are counted; pages it reads without crediting are excluded, since only one engine exposes them.

Every engine was queried through its API on the fast tier it serves by default. The questions come from live panels dominated by three families:

• leisure travel, on which we measured the intermediation layer;
• local professional directories;
• business software discovery.

They are asked in six languages.

One more thing

Four more sources per Perplexity answer does not mean four more of the same names. Compared question by question, on the same day, two engines share a small part of the sources they name, and that has been true of every window we have measured.

So the room that opened on one engine this month opened onto its own web, not onto the shared one. Which engine your market asks is still the first thing that decides whether your name can appear at all.