Search behavior changed faster in the last 18 months than in the previous decade. AI answers now sit above the fold for most informational queries, zero-click results keep climbing, and yet — our client portfolio grew organic revenue 34% year over year. The traffic didn’t disappear. It moved.

This is our data from $50M+ in managed spend across 60 active clients, and the playbook we’ve rebuilt around it.

Informational queries are gone. Good riddance.

The top of the funnel got absorbed by AI answers — but those clicks never converted well anyway. What still drives clicks, consistently, are queries with commercial and transactional intent: comparisons, pricing, alternatives, reviews. Buyers still want to see the source before they spend.

The pattern in the data is unusually clean. Across the portfolio, sessions from queries we classify as informational fell by roughly half. Sessions from commercial-intent queries were flat to slightly up. Revenue from organic rose. If you were measuring sessions, the last eighteen months looked like a catastrophe. If you were measuring revenue, they looked like a mild improvement in efficiency.

That divergence is the whole story, and it explains why the industry’s mood and the industry’s results have been so badly out of step.

What the traffic decline actually consists of

When we break the losses down by page type, they concentrate almost entirely in three categories.

  • Definitional content. “What is X” pages. These were the backbone of a decade of content marketing and they are now answered above the results, in full, without a click. There is no strategy that recovers this traffic, because the user’s need is genuinely met.
  • Simple how-to content. Steps that fit in a paragraph. Same mechanism, same outcome.
  • Aggregated list posts. “Ten best tools for X” written without original testing. AI systems synthesise these trivially, and the synthesis is often better than any individual post.

What did not decline: pages with original data, pages with a point of view that can be attributed to a named person, pricing and comparison pages, and anything requiring an interactive tool. The common factor is that none of them can be summarised without losing the thing that made them worth reading.

Zero-click is not the same as zero-value

The most expensive analytical mistake we see is treating an impression without a click as a wasted impression. In an AI answer, being the cited source is a brand impression delivered at the exact moment of research, to a person who is actively investigating your category.

We started tracking citation share — how often a client is named or linked in AI answers for their priority queries — roughly a year ago. For the clients where citation share rose, branded search volume rose within two quarters, almost without exception. For the clients where it fell, branded search flattened.

The click was never the value. The click was a proxy for the value, and it was a good proxy for twenty years. It is now a partial one, and treating it as complete leads directly to defunding the work that is still producing revenue.

Where we’re winning right now

  • Comparison and alternatives pages — the highest-converting template in our portfolio, and AI engines cite them constantly.
  • Original data studies — proprietary numbers earn links and AI citations simultaneously.
  • Brand search cultivation — social and PR investment now shows up directly in branded query volume.
  • Interactive tools — calculators, configurators and assessments cannot be summarised away, and they collect first-party data as a side effect.
  • Named expertise — content attributed to a real person with a real track record is cited more often and converts better. The bylineless corporate blog post is the format with the least future.

Stop optimizing for the click you lost. Optimize for the citation you can win — AI engines are the new page one.

— Maya Bennett, Head of SEO at Marrky

How AI systems actually choose sources

We cannot see inside the ranking systems, and anyone claiming otherwise is selling something. What we can do is observe which pages get cited across a large sample and look for what they have in common. Four properties recur often enough that we now treat them as working rules.

Extractable structure. Pages that answer a question in a clearly delimited section get cited more than pages that bury the same answer in a narrative. This is not a return to keyword density — it is closer to good editing. A clear heading followed by a direct answer, then the nuance.

Specificity that cannot be paraphrased. A number, a date, a named methodology. Generic advice is synthesised from a dozen sources and credited to none of them. A statistic with a stated sample size gets attributed, because attribution is how the system signals that the claim is not its own.

Corroboration. Claims that appear on several independent, credible sites are treated more confidently than claims appearing once. This makes digital PR and analyst relations meaningfully more valuable than they were three years ago.

Freshness where freshness matters. For queries with a temporal dimension — pricing, regulations, product capabilities — recently updated pages dominate citations to a degree that surprised us. For evergreen topics, publication date matters much less.

The measurement problem

None of this is straightforward to measure, and the honest position is that our instrumentation lags the change by some distance.

We track four things. Citation share, sampled manually against a fixed list of priority queries each month — tedious, and the only reliable method we have found. Branded search volume, which is the clearest downstream signal that visibility is converting into memory. Direct and dark traffic, watched as a trend rather than a number. And organic revenue, which is the only figure that settles arguments.

We have stopped reporting organic sessions as a headline metric entirely. It caused more bad decisions in the last year than any other number on the dashboard, because it fell for reasons that had nothing to do with performance and every executive reads a falling line as a problem to be solved.

What we tell clients to stop doing

Three practices have gone from marginally useful to actively counterproductive.

Publishing to a volume target. A cadence of eight posts a month, set as a goal in itself, now reliably produces a library of content that is invisible to both search and AI systems while consuming the budget that could have funded one genuinely original piece.

Refreshing old posts by changing the year in the title. This worked for a long time. It now produces pages that are demonstrably stale on inspection, and inspection is exactly what the systems doing the citing are performing.

Optimising for featured snippets as a distinct workstream. The snippet has been largely absorbed into the AI answer. The work is not wasted — clear structure still helps — but treating it as a separate goal with its own reporting is measuring a surface that is disappearing.

Budget: where we moved the money

Across the portfolio, the reallocation has been consistent. Roughly a third of content production budget moved into original research — surveys, aggregate analysis of client data, and small studies that produce a number nobody else has. Another meaningful slice moved into digital PR, specifically to get those numbers cited elsewhere. Content volume fell by about forty percent by count and rose in average cost per piece.

Technical SEO spending stayed flat, which surprises people. The fundamentals did not change: crawlability, speed, structured data and information architecture all still matter, and a site with broken basics will not be cited any more than it will rank.

The uncomfortable part

This shift is harder on small teams than on large ones. Original research requires either data you already own or budget to collect it. Named expertise requires a person willing to be named. Both are easier when there are more of you.

The compensating advantage is focus. A small team publishing four genuinely original pieces a year, each with a number nobody else has, will out-perform a large team publishing eighty derivative ones — and we have client evidence for that on both sides. The scale advantage in content marketing was always partly an artefact of a system that rewarded volume. That system is gone.

The playbook, condensed

Audit which of your pages AI engines already cite. Double down on templates with commercial intent. Publish original data quarterly. And measure branded search growth as a first-class KPI — it’s the clearest signal that your brand, not just your content, is winning the category.

Search isn’t dying. It’s consolidating around trust — and that’s an advantage for brands willing to earn it.

What we expect next

Two things look likely, and we are planning for both without betting the strategy on either.

Attribution will get worse before it gets better. As more of the journey happens inside interfaces that do not pass referrer data, the proportion of revenue that arrives with no traceable source will keep rising. Teams that have not built a survey-based self-reported attribution habit will find themselves unable to explain their best quarters.

And the value of a first-party audience will keep climbing. Email lists, communities, and any channel where you reach people without an intermediary deciding whether to show you are worth substantially more than they were three years ago. The strategic logic of the last decade — rent attention from a platform, convert it on your site — is not broken, but the rent has gone up and the terms are less predictable.

Neither of these is a prediction we would defend strongly. Both are cheap to prepare for, which is the only sensible test for a forecast about a system nobody outside a handful of companies can observe directly.

A note on methodology, since we are asking you to trust numbers

The portfolio figures in this piece come from sixty client accounts that were active for the full eighteen-month period, across nine industries, weighted toward business software and professional services. We excluded accounts that changed domain, underwent a major site migration, or altered their measurement setup during the window, which removed eleven accounts from the sample.

Revenue is measured as closed-won value attributed to an organic first session, imported from each client’s CRM rather than taken from analytics conversions. Citation share is sampled manually against a fixed list of between forty and one hundred and twenty priority queries per account, checked monthly, always from the same market and always logged out.

The obvious limitation: this is a sample of companies that hired an agency and kept it for eighteen months, which selects for a certain level of investment and organisational competence. The direction of the findings has held up against everything else we have seen, but the magnitude almost certainly does not generalise to a business publishing occasionally with no dedicated owner.

Four questions we get every week

Should we block AI crawlers? Almost always no. The clients who blocked them saw citation share fall, branded search flatten, and no measurable offsetting benefit. There are legitimate reasons to block — licensing negotiations, genuinely proprietary content — but doing it as a defensive reflex removes you from the surface where your category is being researched.

Is it worth optimising specifically for one AI product? Not as a separate workstream. The properties that get a page cited are broadly consistent across the systems we have sampled, and they overlap heavily with what has always made a page good. Building a distinct strategy per interface is how teams ended up with a folder of dead AMP pages.

Do backlinks still matter? Yes, and arguably more than in 2023, though the mechanism has shifted. Links still carry ranking weight, and independently, being referenced across multiple credible domains appears to increase the confidence with which a claim is attributed to you. The tactics that work are the unfashionable ones: original data, expert commentary, and being genuinely useful to a journalist on a deadline.

How long before we see movement? In our accounts, citation share responds within one to two months of publishing something genuinely new. Branded search follows at two to three quarters. Revenue follows branded search. Anyone promising a faster cycle is either working in a very small niche or measuring something else.

If you only change three things

Most teams reading this do not have the budget to restructure their content operation this quarter. Three changes carry most of the benefit and cost very little.

First, stop reporting organic sessions as your headline. Replace it with organic revenue and branded search volume. This single change removes the pressure to chase traffic that never converted, and it takes an afternoon.

Second, put a real byline on everything, with a real biography and a real track record. It costs nothing, it is the property most consistently shared by cited pages, and it makes your content better because people write more carefully when their name is on it.

Third, publish one thing a quarter that contains a number nobody else has. It does not need to be a large study. An analysis of your own customer data, a survey of two hundred people in your category, a teardown of fifty competitor pages — anything that makes your page the origin of a fact rather than a restatement of one.

Those three changes are what separated the accounts in our portfolio that grew from the ones that did not. Everything else in this article is refinement on top of them.