How do we keep this research current?

The problem: this subject rewrites itself while you're reading about it

Most content about AI search ages the way milk does, not the way wine does. Here's what happened in just the months around this page being written.

In February 2026, Anthropic rewrote its crawler documentation. Where site owners had previously dealt with one declared crawler, there were now three, each controllable separately, each with different consequences for whether your business appears in Claude's answers. Every article on the web describing the old single-crawler picture became quietly, confidently wrong overnight, and most of them are still up, still ranking, still being read.

Around the same time, the most-cited statistic in this field moved. SparkToro's research with Datos had put zero-click Google searches at roughly six in ten in 2024. Its 2026 analysis found 68% of US searches ending without a click in the first four months of the year. Anyone still quoting "60%" is describing a world two years gone, in a field where two years is a geological age.

Neither change announced itself to the people relying on the old information. That's the nature of the problem: the engines and the evidence move, and static content doesn't know it's been left behind. A page can be perfectly researched on the day it publishes and misleading by Christmas. Most publishers handle this by publishing a new article and leaving the old one up to rot, which is good for their page counts and bad for everyone who lands on the stale version. We think that's the wrong trade, so we run a different system. It has five working parts.

Part one: primary sources, named and linked

Every factual claim in our research traces to a primary source: the platform's own documentation, the original study, the official announcement. Not a blog post summarising the documentation. Not a statistics roundup quoting a study three links removed from the data. The thing itself.

And we link it. That sounds trivial; it isn't universal. Plenty of publishers name sources without linking them, sometimes from carelessness, sometimes from a misplaced fear that sending readers to the source makes the article look redundant. Our view is the opposite. If our page adds nothing beyond its sources, a hidden link won't disguise that. What our pages add is the layer the sources can't: what we see in the actual answers engines give, what our client work shows, and what any of it means for a UK business deciding where to spend. The sources carry the mechanics; we carry the judgement. Linking the mechanics is what makes the judgement checkable, and checkable is the only kind of trustworthy that counts.

There's one honest exception, and it's stated when it happens: if we can't verify a source's address is live and correct, the source stays named but unlinked until we can. A wrong link is worse than no link.

Part two: dates that tell the truth

Every source in our references carries a date, and the date means something specific. A study shows its own publication date, so you know how old the finding is, alongside the date we last confirmed it still stands. Living documentation, the platform docs that change without notice, shows only a verification date, because docs pages have no meaningful publication date and pretending otherwise is false precision dressed as rigour.

The same honesty applies to our own pages. Each one declares when it was last reviewed, and that date moves only when a genuine review happened. It would be easy to bump dates on a schedule to look fresh, and some publishers do exactly that. We treat it as lying to the reader, and to the engines, which increasingly reward visible recency and will eventually get better at spotting the fake kind.

Part three: our own numbers are locked before they're published

The client results quoted across this site, the revenue figures, the click growth, the rankings, run on a stricter system still. Every figure lives in a verified internal inventory, traced to production analytics exports, and appears in copy exactly as exported, character for character. A number that hasn't been verified doesn't appear at all: it's held back, visibly gated, until the real data lands. Our flagship sector report is being held unpublished right now for precisely this reason; the structure is built, and it waits on verified numbers rather than launching on plausible ones.

The rule underneath is absolute: no invented figures, no "illustrative" statistics, no rounding a result upward because it reads better. In a field this noisy, the discipline of only publishing what we can prove is not a limitation. It's the product.

Part four: human checks on what the engines actually say

A lot of our research concerns what AI engines do: who they cite, who they recommend, how their answers change. There's a growing industry of tools claiming to measure this automatically, and our experience is blunt: the engines change faster than the tools tracking them, so automated scores tend to describe a version of AI search that no longer exists.

So we check by hand. We ask the engines the questions real buyers ask, in plain language, and record what actually comes back: who gets named, who gets cited, what changed since last time. It's slower than a dashboard and it's the only measurement that reflects what your customers actually experience. Where our guides describe engine behaviour that the platforms haven't publicly documented, we say so in those words, "observed behaviour", so you always know whether a claim rests on the platform's documentation or on our fieldwork. Both are useful. They are not the same thing, and research that blurs them isn't research.

Part five: the quarterly review

Everything above governs how a page gets written. This part governs what happens after, which is where most publishers stop trying.

Every three months, every page in this section goes through a source review. Every outbound link is fetched and confirmed live and still pointing at what we said it points at. Every sourced claim is re-read against its primary source as that source stands today, not as it stood when we wrote. Each engine gets a check for material changes since the last pass: new crawlers, renamed surfaces, superseding research. Whatever moved gets corrected, on the same page, at the same address, with the dates updated to say so.

That last part matters more than it looks. When something changes, we don't publish a new article and abandon the old one to mislead whoever finds it. The page you're reading is the page we maintain: same address, corrected substance, honest dates. Bookmark a guide here and it stays the current answer, or it tells you plainly when it was last checked so you can judge for yourself. The February 2026 crawler change is our standing reminder of why: it was caught and corrected in our Claude guide within the same quarter, and the review exists so that kind of catch is systematic rather than lucky.

When we get something wrong

We will, eventually, despite all of the above; anyone publishing at volume in a moving field who claims otherwise is selling something. When it happens, the correction goes on the page itself, the dates change to reflect it, and if the error was material we say what changed rather than silently patching it. The willingness to be visibly corrected is most of what separates a source from a content farm.

What this means for you

Practically, three things. If you cite our research, the dates tell you exactly how current every claim is, and the links let you check us in one click. If you're comparing what you read here against something elsewhere that says different, check the dates and the sources first; in this field, most contradictions are just timestamps in disguise. And if you want the same discipline pointed at your own visibility, the AI-search visibility check is the working version of part four, run on your business, free.