You can hold position one and still not appear in the answer above it.
That sentence is the entire difference between traditional search and AI Mode, and it is the part that breaks reporting. Your rank tracker says you won. The customer never saw you, because the assistant built an answer from three other sources and the customer stopped reading.
Selectability is the probability that a generative system picks your content for the answer it is assembling. Ranking is your position in a list of links. They are related, and they are no longer the same thing.
This guide is strictly about that distinction inside Google AI Mode: what actually changed, what stubbornly did not, and how to measure the difference without fooling yourself. If you want the wider comparison of SEO against AEO, GEO and the agent layer, that is a different question and we answer it in our AI optimization explained guide.
The Evidence That They Have Come Apart
The clearest public measurement of this gap comes from Ahrefs, which found that only 38% of AI Overview citations came from pages ranking in the top 10, down from 76% (Ahrefs).
Sit with that number for a moment. It means the majority of citations now go to pages that are not in the classic top ten for the query. Ranking still correlates with being cited, but it has stopped being a reliable proxy for it. A brand can be strong on one scoreboard and absent from the other.
There is a second measurement worth knowing: these answers are unstable. One study found that repeating identical prompts rarely returns the same set of brands (Search Engine Land). That is not a bug you can optimise away, it is a property of the system, and it has consequences for how you report.
Branded behaviour is shifting too, with AI Overviews appearing more often on branded queries (Search Engine Land). Your own brand terms are no longer a guaranteed clean path to your site.
What Google Actually Says
Before anyone sells you a selectability methodology, read what the platform states in writing.
- There is no special optimization for AI features beyond ordinary good search practice (Google).
- Structured data is not a requirement to appear in AI features (Google).
- The guidance on helpful, people first content remains the stated basis for what surfaces (Google).
So we have an apparent contradiction. The data shows ranking and citation diverging. The platform says there is no separate optimization.
Both are true, and resolving that is the useful part of this article. Google is saying there is no secret lever. The data is saying the thing being rewarded is not the same thing rankings measure. The work is still SEO; the target within it has moved.
Why the Mechanics Produce Different Winners
A ranked list and a generated answer are built differently, and the differences explain the divergence without needing any secret ranking factor.
One query becomes many
A conversational question gets decomposed into several underlying queries, and the answer is assembled from results across all of them. Your page is not competing for one keyword. It is competing to be the best source for one component of a larger answer, and a page that half covers ten things loses to a page that fully covers one.
The unit is a passage, not a page
A ranked result rewards the page. A generated answer pulls the specific passage that resolves a sub question. A page can rank well on aggregate strength while containing no passage that cleanly answers the precise thing being asked.
There are fewer slots, and they are not ordered
A results page has ten positions. An answer has a handful of citations with no clear hierarchy. The distribution is harsher: being eleventh in a ranked list still earns some traffic, while being the fourth best source for an answer that cites three earns none.
Confidence shapes selection
A system assembling an answer it will present as fact behaves conservatively. Claims that are specific, attributable and corroborated elsewhere are safer to repeat than confident marketing assertions that appear nowhere else.
What Changes and What Does Not
| Traditional search | AI Mode | |
|---|---|---|
| What the user sees | A list of links | An assembled answer with a few citations |
| Unit of competition | The page, for a query | The passage, for a sub question |
| Number of winners | Ten positions, ordered | A handful of sources, unordered |
| How the user arrives | Clicks through | May never click |
| Measurement | Position, clicks, impressions | Citation presence, share of voice |
| Stability | Reasonably stable | Varies between identical runs |
| What wins | Relevance plus authority | Relevance, specificity, extractability, corroboration |
What does not change, and this is most of it:
- The page must be crawlable, renderable and indexable. Nothing happens otherwise.
- It must be accurate. Errors are costlier when a machine repeats them as fact.
- It must be genuinely useful to a person, which remains the stated basis for what surfaces.
- It must load and function, on the devices and connections your audience actually uses.
If a provider’s AI Mode proposal does not spend most of its budget on those four, ask why.
What Genuinely Improves Selectability
Four areas, in the order we would fund them.
1. Cover the sub questions, not just the topic
If a question decomposes into six parts, a page that answers all six is a better source than six pages answering one each, and far better than one page that gestures at the topic. Audit your key pages against the questions people actually ask, and fill the gaps in place rather than publishing new thin pages. Our AEO and why it matters guide covers the question mapping method.
2. Make passages extractable
Put the answer first, then the explanation. Use a heading that matches the question. Keep the resolving statement in a single self contained paragraph that makes sense lifted out of context, because that is exactly what will happen to it. Tables and lists carry comparative and sequential information more reliably than prose.
3. Be specific and attributable
Numbers with sources beat adjectives. Dates beat “recently”. Named examples beat “many clients”. A specific, sourced claim is safer for a system to repeat, and safety is doing more work here than persuasion.
4. Get corroborated elsewhere
Citation studies consistently show community, professional and review platforms heavily represented among cited sources (Search Engine Land). What others say about you feeds whether you are selected, which is why purely on site work plateaus.
On structured data: add it because it makes facts unambiguous and it must match the visible page (Google), not because it is a requirement for AI features. It is not one.
A Worked Example: One Page, Two Scoreboards
A business software company has a page targeting a competitive category term. It holds position three. Organic traffic is healthy. The marketing team reports success.
Then someone asks an assistant the question a buyer would actually ask: which tool is best for a team of fifteen that needs a specific integration and has a modest budget. The answer names four competitors. Not them.
What went wrong, in order:
- The question decomposed and the page did not. The real question contained three constraints: team size, a specific integration, and budget. The page addressed the category broadly and none of the three directly.
- There was no extractable passage. Pricing sat in a table on a different page. Integration detail was a logo grid, which is an image. Team size suitability was implied by tone rather than stated.
- The claims were unquotable. “Powerful collaboration for modern teams” cannot be repeated by a system presenting an answer as fact. “Suitable for teams of 10 to 50, with native integration for X” can.
- There was no corroboration. Competitors appeared on review platforms and in community threads. This company had a listing nobody maintained and no community presence.
None of those four failures affects ranking. The page ranks third precisely because classical signals are fine. That is the divergence in a single example, and the fixes are all content and evidence work rather than anything exotic.
Three Claims About AI Mode SEO That Are Wrong
“You need to chunk your content for AI”
Google’s guidance explicitly lists breaking content into small pieces for AI as unnecessary (Google). Writing clearly, with headings that match questions and self contained answers, achieves the actual goal. Mechanically fragmenting your pages does not, and it tends to produce worse content for humans.
“Schema is how you get into AI Mode”
Structured data is not a requirement for AI features, and a large schema project is not a route to citation. Use markup to state facts unambiguously about entities, products and offers, make it match the visible page, and spend the rest of the budget on coverage and corroboration.
“Rankings no longer matter”
Rankings still correlate with citation, still drive a large share of discovery, and still describe how most people find most things. The honest version of the claim is narrower: rankings alone now overstate your visibility, so they need a second measurement alongside them rather than replacing them.
How to Measure Selectability Without Fooling Yourself
This is where most reporting in this space falls apart, usually in one of three ways: a single flattering screenshot, a vendor score with no stated method, or week on week tracking of a metric too volatile to read weekly.
A defensible method:
- Build a fixed panel of twenty questions your buyers genuinely ask, in their words.
- Run each one at least three times, in a clean session, recording every brand named and every source cited.
- Calculate two numbers: how often you are cited, and your share of brand mentions against the competitors you actually lose deals to.
- Repeat monthly, never weekly, against the same unchanged panel.
- Report the trend with the variance visible, not a single figure that implies a precision the data does not support.
The repetition matters because of the instability finding above (Search Engine Land). One run is an anecdote. Three runs across twenty questions, repeated monthly, is a trend you can act on.
Keep classical reporting alongside it. Rankings and organic traffic still describe a large share of how people find you, and our complete SEO checklist remains the baseline that everything here sits on top of.
So Is SEO Dead?
No, and the question is usually asked by people selling the replacement.
What has actually happened is narrower and more interesting. The scoreboard that described your visibility for twenty years now describes part of it. Rankings still matter, still drive traffic, and still correlate with citation. They have simply stopped being the whole picture, which means reporting that stops at rankings now overstates how well you are doing.
The practical response is not to abandon anything. It is to keep doing the work, add a second measurement, and change what you optimise toward within the same discipline: specific sub questions rather than broad topics, extractable passages rather than long preambles, sourced claims rather than assertions, and independent corroboration rather than louder self description. For the citation layer specifically, our GEO definitive guide goes deeper.
Where YuvGro Fits
We report on both scoreboards, because reporting on one is how brands end up confident and invisible. You get classical rankings and organic performance alongside a selectability baseline: a fixed question panel, repeated runs, citation rate and share of voice against your real competitors, tracked monthly with the variance shown.
That work sits inside our AEO services.
Frequently Asked Questions
What is selectability in SEO?
Selectability is the probability that a generative system chooses your content as a source for the answer it assembles, rather than the position your page holds in a list of links. It matters because the two have come apart: Ahrefs found that only 38% of AI Overview citations came from pages ranking in the top 10, down from 76%. A page can rank well and still never be selected, which is why rank tracking alone now overstates visibility.
What is the difference between ranking and being cited in AI Mode?
Ranking is your ordered position among ten links for a query. Being cited means a generated answer used your content as a source, usually for one specific part of a larger question. The unit differs too: ranking rewards the page while citation pulls a particular passage. There are also fewer citation slots and no clear order, so being the fourth best source for an answer that cites three earns nothing.
Is SEO dead because of AI search?
No. Pages still have to be crawlable, accurate, genuinely useful and functional before anything else happens, and Google’s guidance states there is no special optimization for AI features beyond ordinary good practice. What changed is that rankings now describe only part of your visibility. The correct response is to keep the existing work, add a second measurement for citations, and shift emphasis toward specific sub questions and extractable passages.
How do I increase my chances of being cited in AI Mode?
Cover the specific sub questions a topic decomposes into rather than gesturing at the topic broadly. Write so passages can be lifted out and still make sense, with the answer first and a heading matching the question. Be specific and attributable, since sourced numbers and named examples are safer for a system to repeat than adjectives. And build independent corroboration, because what third party sources say about you feeds whether you are selected.
Does structured data help with AI Mode?
It helps machines confirm facts unambiguously, but it is not required. Google states directly that structured data is not a requirement to appear in its AI features. Add it where it describes real entities and offers accurately, make sure it matches what a visitor can actually see on the page as the structured data policies require, and do not treat a large schema project as a route to AI visibility on its own.
How do I measure AI Mode visibility?
Build a fixed panel of about twenty questions your buyers genuinely ask, run each at least three times in a clean session, and record every brand named and source cited. Track two numbers: how often you are cited, and your share of brand mentions against real competitors. Repeat monthly rather than weekly, against an unchanged panel, and report the trend with variance visible. Weekly reporting is meaningless given how much these answers vary between runs.
Why do AI answers change when I ask the same question twice?
Because variability is a property of how these systems generate answers, not a fault you can optimise away. One study found that repeating identical prompts rarely returns the same set of brands. The practical implication is that no single result proves anything, in either direction. Measure with repeated runs across a fixed question set and look at trends over months, and treat a single flattering or alarming screenshot as an anecdote.
Should I still track keyword rankings?
Yes. Rankings still drive a large share of discovery and still correlate with being cited, even though the correlation has weakened. Dropping them would leave you blind to most of your traffic. The change is that rankings alone now overstate your visibility, so they belong alongside a citation measurement rather than being replaced by one. Report both, and expect them to disagree sometimes.
Do AI Overviews appear on branded searches?
Increasingly yes, and that is a meaningful shift. Reporting has shown AI Overviews appearing more often on branded queries, which means even searches for your own name may be answered before anyone reaches your site. The practical consequences are to make sure the facts about your brand are accurate and consistent across the web, and to track your own branded queries in your citation panel rather than assuming they are safe.