Displacing and joining are two different jobs, and one engine asks for each

Take the same 102 questions per market, the same three markets and the same day, and the two engines produce source tables of completely different shapes. Reading taken on 15 October 2026: the leading host on the ChatGPT side turned up in 176 of the 306 readings that engine gave us, and the leading host on the Gemini side turned up in 19 of its 291. One table has an entrenched leader. The other has nobody in front. The work those two situations ask for is not the same work, and almost no published advice separates them.

Against a concentrated table you are displacing something. Against a flat one you are joining a crowd. Different goals, different page, different expectation of how long it takes.

The two shapes, and what each one asks of a publisher

	A concentrated table	A flat table
What the top looks like	One host in a large share of the readings	Several hosts at a handful of citations each
The job	Displace an incumbent that already answers the question	Join a set where nothing has settled the question
What wins it	Being better on the exact question the incumbent answers, or finding questions it does not answer	Being specific and current on a question nobody has answered well
How a result shows up	Rarely, and then it holds	Sooner, and it moves between readings
What to report	Whether the incumbent's share moved	Whether you are in the set at all, and how often

The row that matters most is the last one. A plan measured against a leaderboard on a flat table will look like failure for months while the real result, presence in a shifting set, goes unrecorded because nobody kept a column for it.

Measuring which shape your own category is in

You do not need our numbers for this and you should not use them, because the shape is a property of a category on an engine rather than of the engine alone. The measurement takes an afternoon.

• Take thirty to a hundred buying questions in your buyers' words, with no company name in any of them.
• Run the identical wording on each engine you care about, one engine per column.
• For each engine, count how many of its sourced readings cited its own most cited host.
• Divide that by the number of its sourced readings, and write the two numbers down rather than the ratio alone.
• Do it again in a month. The shape moving is itself a finding.

Two engines in our own first run came out at 176 of 306 and 19 of 291 on that measure. A category where both engines come out flat is a category with room on both. A category where both are concentrated is one where question selection matters more than writing quality.

Why a blended score destroys exactly this information

Average the two readings above and you get a middling figure that describes neither engine. Worse, it hides the only instruction the data contains, which is to do different work on each surface. This is the practical reason we report one column per engine with its own denominator and never publish a single combined visibility score.

The same logic applies to your own reporting. If a tool hands you one number across several apps, ask what the per app numbers were, because a plan cannot be written against an average of two different games.

What this comparison does not say

It is one date, 15 October 2026, one bank of 102 questions per market, and two apps. Perplexity, Claude and Google AI Overviews were asked nothing, and three of the five surfaces our product reads are therefore unmeasured in every figure above. One reading per question means these are tendencies on a day rather than fixed rankings. There is no per industry rate inside the bank, because no question named an industry. And every count is of a domain in a source list, not of a business named in the text of an answer.

We sell the measurement that produces readings like this and the writing that follows it, so our own result from the same run is on the record: 3 readings of 597 cited a page of ours, and none of those were on the flatter engine. The search console side of our own record, with both windows and the export date printed, is our proof page.

The mistake that comes from reading the wrong table

There is a natural assumption worth naming because it is the wrong way round in our own data. A table dominated by one company's documentation looks as though it must belong to that company's own assistant. On this reading it does not. The concentration sits on the other app, and the engine run by the owner of those documentation sites spread its sourced answers thinly across small marketing and vendor pages instead.

Two practical consequences follow from getting that the right way round. First, you cannot infer which engine favours which publisher from who owns what; you have to read it. Second, a tactic justified by whose engine it is rather than by a count is a tactic justified by nothing. Count first, then decide, and keep the count dated so next quarter's version of the argument has evidence instead of intuition.

Questions about concentration and flatness across engines

What were the two top host figures?

176 of 306 readings on one engine and 19 of 291 on the other, both from the same questions on 15 October 2026.

Which shape is easier to enter?

The flat one, because nothing has to be displaced. What it does not promise is that staying in the set is easy, since a set with no leader also has no settled position to hold.

Does a concentrated table mean I should give up on that engine?

It means you pick questions rather than fight the leader. A concentrated table is usually concentrated on one kind of question, and the rest of the bank behaves differently.

Should I report one score across my engines?

One column per engine, each with its own denominator. An average across two tables of different shapes is a number that cannot be acted on in either direction.

Which engines produced these two readings?

ChatGPT and Gemini, asked the same 102 questions per market on 15 October 2026 in the United States, the United Kingdom and India. The full method and source mix are published as 597 readings across ChatGPT and Gemini in three markets.

Decide which job you are doing before you brief the page

Write it at the top of the brief: displacing, or joining. The first needs a page that beats a specific incumbent on a specific question, and a patient reporting line. The second needs volume of specific, current, dated pages and a reporting line that counts presence rather than position. Briefing the second job and measuring it like the first is the most common way a reasonable plan gets cancelled early.