What's Different

1. There Is No Position, Only Inclusion

A results page has ten slots and a long tail below them. Fourth place is a real business. Twentieth place still earns something.

A generated answer naming three sources has three slots and nothing underneath. There is no fourth place and no consolation traffic. This is the difference that makes every other one matter: it converts a gradient into a threshold.

Practically, it means incremental improvement can produce nothing at all for a long time and then produce everything at once. Teams used to watching a ranking climb steadily find this disorienting, and some abandon work that was about to succeed.

2. Value Arrives Without A Click

When your comparison table is summarised into two sentences and credited to you, you have been recommended to someone who asked. Your analytics record nothing.

This breaks the measurement model rather than the value model. Sessions were always a proxy for influence, and they are now a worse proxy than they were. The response is not to dismiss what you cannot count. It is to count it differently, through mention frequency and share of voice across platforms, alongside the traffic you can still see.

3. Selection Discards You Before Judging You

Retrieval finds more material than fits in the context window, so most candidates are cut before merit is assessed. In classic search, a comprehensive page competed on its merits. Under retrieval, a comprehensive page can lose to a narrower one purely because its answer is easier to locate and extract.

This inverts a decade of content advice. Length was a proxy for depth and depth was rewarded, so length was rewarded. Depth still wins. Length without depth now actively costs you, because every padded paragraph makes the useful passage harder to isolate.

4. The Query Is Rewritten Before It Is Searched

The user asks one thing. The system decomposes it and searches for something else, in its own words, often several times.

Optimising for exact phrasing therefore matters far less than it did. What matters more is whether the system already associates you with the category it will search for. If it does not, the query it generates will not surface you and your page quality is irrelevant to the outcome.

This is the change that moves the most work off your own website. Being described as part of your category, by other people, in places these systems read, is now a retrieval prerequisite rather than a branding nicety.

5. Results Vary Between Identical Runs

Ask the same question twice and get two different answers. This comes from sampling in the model, variation in the searches it generates, and any memory it carries.

Rank tracking assumed a stable answer that could be checked. Nothing here is stable in that sense. A single observation is an anecdote, and treating it as a reading is the most common measurement error in the field.

The working replacement is frequency: run each prompt several times in fresh sessions and record appearances out of runs. Three out of five is a number that means something. "I checked and we were there" is not.

6. The Platforms Disagree About Which Sources Exist

Google and Bing broadly agreed about which pages were relevant. The assistants do not. In the citations analysed by the 5W Public Relations AI Platform Citation Source Index published in May 2026, only about 11% of domains cited by ChatGPT were also cited by Perplexity.

That figure comes from an agency synthesis rather than peer-reviewed work, so treat the precision cautiously. The magnitude is what matters, and it is consistent across the underlying studies: platform-specific measurement is not a refinement, it is the minimum viable measurement. A blended score across systems that disagree this much destroys the information you needed.

Side By Side

Dimension	Search	Generated answers
Unit of success	Position in a ranked list	Inclusion, with no second page
How value arrives	A click	A click, a citation, or an unlinked mention
Effect of length	Neutral to positive	Costly when it dilutes the answer
Query matching	Against the user's words	Against the system's rewritten query
Stability	Same query, similar results	Same query, varying results
Cross-platform agreement	High	Low, roughly 11% domain overlap in one large synthesis
Right measurement	Position and sessions	Frequency across runs, per platform, plus mentions

Translate This Into Your Own Reporting

The six differences are only useful once they change something you do. This takes an afternoon:

1. Add mention tracking alongside session tracking for your ten most commercially important questions. Log all four outcomes, not just linked citations.
2. Replace single checks with frequency: three runs per prompt, fresh sessions, recorded as appearances out of three.
3. Split every report by platform. Never average across them.
4. Take your three longest important pages and find the passage that answers the question. If you cannot find one in ten seconds, neither can a selection step.
5. List the category phrases the assistants use when they search on your behalf, and check whether independent sources describe you using those phrases. Where they do not, that is your off-site work queue.

Key Takeaways

• Inclusion replaces position, which turns gradual progress into a threshold effect.
• Value now arrives without a click, so reporting built only on sessions will show working strategies as failures.
• Selection discards material before merit is judged, which is why padding costs what it never used to.
• Systems search their own rewritten queries, moving decisive work off your website and onto how others describe you.
• Answers vary between runs. Measure frequency across repeated fresh sessions, per platform, never blended.
• Roughly 11% domain overlap between two major platforms makes per-platform measurement the minimum, not an upgrade.