The reporting pattern has become familiar. Impressions are flat or up. Clicks are down twenty to forty percent. Rankings have not moved. Somebody in the meeting says the algorithm changed, somebody else says the site got penalised, and a quarter goes into fixing something that was never broken.
Nothing is broken. The traffic went somewhere specific, and it is worth being precise about where — because the correct response depends entirely on getting the diagnosis right.
What genuinely changed
The answer moved to the results page. For a large class of queries, the user no longer needs to visit anything. The system composes a response and cites a handful of sources. You can be the top result and the best source and still receive no visit, because the visit stopped being necessary.
Selection is not ranking. In classic search, position one gets the click. In an answer surface, a set of candidate documents gets retrieved and a subset gets cited. Being retrieved is necessary. Being cited is a second, different decision — made on how quotable your passage is, not on where you sat in a list.
The unit of value shrank from the page to the passage. A model does not cite your article. It cites a sentence or two from it. A superb 3,000-word guide whose key claims only make sense in context is close to uncitable. Three clean standalone sentences from a mediocre page will be quoted every time.
Brand mention started to matter independently of links. These systems build associations from how entities are described across the web, not only from hyperlinks. Being consistently named as a provider of a thing is now a retrieval signal in its own right.
What did not change at all
This is the part that gets lost, and it is the more useful half.
Retrieval still runs on conventional indexes. ChatGPT search, Perplexity, AI Overviews and Copilot are all pulling from web indexes built by conventional crawlers. If you are not crawlable, not indexed, or not relevant, you are not in the candidate set. There is no separate AI index you can optimise your way into while ignoring the normal one.
Authority still decides eligibility. The systems disproportionately cite sources that were already well-regarded. There is no shortcut where a thin site outranks an authoritative one because its formatting is tidier.
Technical fundamentals still gate everything. Crawl access, render, indexation, site structure, internal linking. All of it still applies, unchanged. This is why I still start SEO engagements with a crawl and an index reconciliation rather than with content.
Commercial intent still converts. Someone comparing two providers, checking a price, or looking for a supplier in their city still wants to see the actual thing. Answer surfaces absorb informational traffic far more than commercial traffic.
The mental model that holds up: AI search changed the presentation layer and the selection layer. It did not change the retrieval layer. Optimise for all three or you will be tuning a passage nobody can fetch.
The asymmetry nobody plans for
Not all traffic loss is equal, and treating it as one number hides the only thing worth knowing.
Rough shape of what happens, by query type:
- Definitional and simple informational — heaviest losses. "What is X" is fully resolved by a synthesised paragraph. There is nothing left to click.
- How-to and procedural — substantial losses for short procedures, much less for anything requiring depth, screenshots, or tools.
- Comparison — modest. Buyers want to see the actual options, and they distrust a summary of a decision they are about to pay for.
- Commercial and transactional — least affected. Nobody buys from a paragraph.
- Branded — essentially unaffected. Someone searching your name wants you.
If your losses are concentrated in the first two rows, your traffic mix was informational and the revenue impact is probably far smaller than the session count suggests. If you are losing commercial and comparison clicks, that is a genuinely different problem and it is probably not AI search.
Segment before you react. I have watched teams rebuild an entire content programme over a decline that was almost entirely definitional glossary traffic that never converted once.
What to actually do
Segment the loss before you touch anything
In Search Console, split queries by intent and compare clicks and impressions period over period for each segment. You are looking for where impressions held and clicks fell — that gap is the answer-surface effect. Where both fell, you have an ordinary ranking or demand problem, and AI is a red herring.
Do this first. Everything after it depends on knowing which problem you have.
Make your key claims survive extraction
The practical test: take any important sentence out of your page, put it in isolation, and read it. If it needs the previous paragraph to make sense, it will not be quoted.
That means resolving pronouns into nouns, putting the conclusion before the reasoning, and keeping load-bearing claims to roughly the length of a single complete thought. It is a real rewrite, not a formatting pass, and it is the bulk of what GEO and AEO work actually consists of.
Give the model something attributable
Cited content tends to carry something concrete: a figure with a stated method, a named source, a date. Vague competence does not get quoted because there is nothing to attribute. If you have proprietary data, publish it with its methodology attached — that is the most reliably citable asset most companies already own and never use.
Fix entity consistency
If your company is described four different ways across your site, LinkedIn, and directories, the systems have four weak associations instead of one strong one. Pick one description of what you do and use it everywhere, verbatim. This is unglamorous and it moves the needle more than most content work.
Change what you report
Sessions are becoming a worse proxy for demand every quarter. More durable measures:
- Citation rate on a fixed prompt set, re-run on a schedule
- Branded search volume, which rises when people encounter you in answers and go looking
- Direct and dark traffic, the usual destination of an unattributed AI referral
- Conversion rate of remaining organic, which typically rises — the people still clicking through are the ones who needed more than a summary
That last one is worth sitting with. Losing the visitors who only wanted a definition is not a loss. If your leads held steady while sessions fell thirty percent, the funnel got more efficient, and reporting it as a decline is a reporting failure rather than a marketing one.
The honest uncertainty
Anyone giving you precise numbers about AI citation behaviour is overstating what is knowable. Selection is non-deterministic, it varies by prompt phrasing, it changes with every model update, and none of the platforms report it. Two identical prompts an hour apart can cite different sources.
What is knowable, and worth working on:
- Whether crawlers can reach you — directly testable
- Whether your content survives being extracted — readable by inspection
- Whether your claims carry attributable evidence — a content decision
- Whether your entity is described consistently — an audit
- Whether your citation rate on a fixed prompt set moves — measurable over time
That is a real programme of work with verifiable inputs. It is less satisfying than a number that says you gained 43% AI visibility, and it has the advantage of being true.
If clicks are falling and you cannot tell which of these applies to you, send me the URL. Segmenting the loss is the first thing I do, and it usually changes what the problem turns out to be.