The number that matters is not the ranking
If you track rankings only, an AI Overview is invisible. Your position is unchanged, your impressions may even rise, and your clicks fall. Every dashboard says nothing is wrong while the traffic line goes down.
The pattern to watch for is divergence: impressions holding or climbing while clicks drop. That gap is the room an AI-generated answer occupies. It is not proof — a title change produces something similar — but it is the cheapest signal available and it appears weeks before a revenue conversation.
Three things worth measuring, not one
Clicks and impressions, separately. Not a blended “traffic” figure. The ratio between them is where the AI Overview shows up.
AI citations. Whether an answer engine names you at all, and how often. This is a different question from ranking, and it has a different answer — plenty of pages are cited by ChatGPT that never sit in the top three.
Citation rate, with a sample size. This is the part almost everyone gets wrong. AI answers are
non-deterministic; running one query once and recording “cited” or “not cited” is not a
measurement. Run each query several times and report a rate: cited 3/5. Then compare rates week
over week, not single observations.
What to do about it
The reflex is to chase the answer box. That mostly does not work, for a structural reason: AI answers are assembled from sources chosen for extractability and consensus, not position.
What does work is duller:
- Make key claims self-contained. A paragraph that only makes sense after the two before it cannot be lifted, so it will not be.
- Lead with the answer. The first sentence of a section is the part that gets used.
- Put numbers in, with their source and date. Cited statistics raise visibility substantially in controlled studies; unsourced adjectives do nothing.
- Be present off your own domain. A large share of citations come from third-party sources — review platforms, communities, publications. Your site alone is not the surface being read.
The honest position
Nobody can currently promise you AI Overview traffic. The behaviour changes without notice; retrieval updates have moved citation patterns overnight for reasons unrelated to anyone’s work. What you can do is measure the divergence early, keep the structure extractable, and not confuse “our ranking is fine” with “we are visible.” Those stopped being the same claim a while ago.