Guide · · 6 min read
AI search changed SEO. What actually matters now
When the answer appears above the links, ranking first stops being the goal. What to measure instead, and what has not changed at all.
By Manish
For twenty years the job was to rank. Reach position one, collect the click, repeat. That model assumed the user would click something, and increasingly they do not, because the answer is already sitting on the results page.
This has produced a great deal of panic and a smaller amount of useful thinking. Here is the useful part, along with the parts that have not changed at all and are being ignored in the noise.
What actually changed
Three things, and they compound.
Answers moved above the links. A generated summary sits where the first result used to be, and for informational queries it frequently satisfies the user completely. The click that funded the page never happens.
Assistants replaced some searches entirely. A meaningful share of questions that would have gone to a search engine now go to a chat interface, which returns an answer with a handful of citations rather than ten options.
Citation replaced ranking as the unit of visibility. Being the source quoted in an answer is now worth more than being the fourth blue link, and those two things are not correlated as tightly as you would expect.
The metric moved
Position is no longer the thing to watch, and continuing to report on it produces a picture that looks stable while the underlying reality shifts.
A page can rank fourth and be the source an assistant quotes. Another can rank first and never be mentioned. If your dashboard only tracks position, both of those look identical and neither is.
- Track citation share: how often you appear in generated answers for the questions you care about.
- Track it per question rather than per keyword, because assistants answer questions and keywords are an artefact of the old interface.
- Track who is cited alongside you. Appearing with three competitors is a different position from appearing alone.
- Keep tracking clicks, because they still convert, and the traffic that does arrive is more qualified than it used to be.
A category of tooling has appeared to measure exactly this. It is young and the methodologies vary, so treat the absolute numbers cautiously and watch the trend.
What gets cited
Assistants cite sources that are easy to verify and hard to dispute. In practice that means specifics, and specificity is precisely what a decade of SEO advice trained people out of.
- Numbers with dates attached. Published in March 2026 beats recently, always.
- Stated methodology. How you know something is what makes the claim quotable rather than assertable.
- Narrow claims. A page that says one thing precisely gets cited for that thing. A page that says twenty things vaguely gets cited for none.
- Original data. Anything only you have is the strongest citation magnet available, because there is no alternative source.
- Clear structure. Headings that match real questions make the relevant passage easy to locate and extract.
The vague, keyword-padded page that ranked well in 2020 is exactly the page a model has no reason to quote, because it contains no sentence worth quoting.
Write the sentence you would want quoted. Then make sure it is defensible, because being quoted and wrong is considerably worse than being ignored.
Which pages are actually at risk
The damage is uneven, and treating it as uniform leads to panicked rewrites of pages that were never in danger.
Most exposed are simple informational queries with a single factual answer. What is the capital of, how many grams in, what does this acronym stand for. Those clicks are gone and are not returning.
Least exposed is anything requiring judgement, comparison across many options, current data, or a transaction. People still click when they are deciding something that matters, because a summary is not sufficient basis for spending money.
Which means directories, comparison content, original research and anything transactional are in a stronger position than the panic suggests. Thin informational content is in a much weaker one, and it was always the least defensible thing on the internet.
What has not changed
Being genuinely useful still wins, and it wins for the same reason it always did.
Sites that answer the question completely get linked, cited and returned to. Sites that pad the answer to hit a word count get skipped by readers and, now, by models that have no reason to quote filler. That was true when the ranking was ten blue links and it is true when the ranking is one paragraph.
Technical fundamentals still matter too. If a crawler cannot read your page, an assistant cannot cite it. Clean markup, fast loads, working structured data and content that exists in the HTML rather than only after JavaScript runs. None of that is new and all of it is now more consequential.
What to actually do
Audit your pages against a simple question: does this contain a sentence somebody would quote? If not, that page has no future in either ranking system, and adding keywords will not give it one.
Add specificity to the pages worth saving. Dates, numbers, methodology, and a clear statement of what you know and how. This is more work than keyword optimisation and it is the work that transfers.
Invest in whatever only you can produce. Original data, genuine testing, a real point of view. That was always the strongest position and it is now close to the only defensible one.
And stop optimising for a machine you cannot see. The tactics changed and the underlying trade did not. Say something true, say it specifically, and be the page somebody would want to point at.
Structured data matters more than it did
Schema markup was always worth doing and easy to deprioritise. It is now closer to essential, because it is the most reliable way to tell a machine what your page actually asserts.
An assistant extracting a claim from prose has to interpret. An assistant reading structured data does not. Marking up products, reviews, FAQs, articles and organisations removes ambiguity at exactly the moment ambiguity costs you a citation.
Keep it honest, and this matters more than it used to. Marking up ratings you do not have, or reviews you generated, is the fastest way to have your structured data ignored across the board. Machines cross-check, and a source caught overstating once is discounted permanently.
The freshness signal got stronger
Assistants prefer recent sources for anything that could have changed, and they are more aggressive about it than search rankings ever were.
That makes visible dates important. A page with a clear published and updated date is easier to trust than one with no date at all, and undated pages get quietly deprioritised for anything time-sensitive.
It also makes maintenance a genuine strategy rather than an afterthought. Updating an existing strong page with current data frequently returns more than writing a new one, and almost nobody allocates time to it because new content is easier to justify.
Traffic will fall and convert better
The pattern reported consistently by sites through this transition: total sessions down, conversion rate up, absolute conversions roughly flat or slightly ahead.
The explanation is straightforward. The visitors you lost were the ones who wanted a fact and left. The ones still arriving have already had the summary and clicked anyway, which means they want depth, or they are comparing, or they are about to act.
This makes traffic a worse metric than it already was. If your reporting leads with sessions, this transition will look like a catastrophe while the business underneath is fine. Lead with conversions and the picture is accurate.
Do not chase the mechanism
There is a growing industry of tactics claiming to game AI citation. Specific formatting, particular phrasings, structural tricks that supposedly make a model more likely to quote you.
Some of it works briefly. All of it is the same trade as keyword stuffing was, which is borrowing short-term visibility against the risk of being discounted when the mechanism changes. The mechanisms here change considerably faster than search algorithms ever did.
The durable position is unchanged and unexciting. Publish things that are true, specific, current and hard to get elsewhere. That worked for ten blue links, it works for generated answers, and it will work for whatever replaces them, because every one of those systems is trying to find the same thing and is only getting better at it.
Find the right SEO tool
Compare the SEO tools we have reviewed, with prices and honest pros and cons.