The domain fingerprint doesn’t survive the transition

Duane Forrester wrote a piece this week that has been sitting in my head for two days. He calls it the fingerprint question: every domain you've ever optimised carries a granular record of that work — link profile, internal architecture, Core Web Vitals, E-E-A-T signals, schema coverage, crawl history, the lot. That record is real. It's what the industry has been shaping for fifteen years.

His question is whether that record comes along when search stops being the destination and becomes the substrate under an answer engine.

He answers it partially. For Google, the inheritance is architectural — AI Mode and AI Overviews are rooted in the same ranking systems that produce the blue links, so whatever your domain was worth in ranking terms is what feeds the generated answer. He's right. Google has said this out loud in its own documentation. That's the easy half.

The hard half is what happens when the answer isn't Google. And the hard half is where the industry is quietly getting something important wrong.

The fingerprint is a Google artefact. That's the whole problem.

Everything on Forrester's list — link velocity, referring-domain diversity, anchor-text distribution, click depth, Core Web Vitals, E-E-A-T operationalisation via Search Quality Rater Guidelines, crawl cadence — is a Google-shaped record. It exists because Google's crawler recorded it and Google's ranking systems weighted it. The fingerprint is not an abstract property of your domain. It's a specific set of measurements taken by a specific vendor, over years, using their own instruments.

Your Google fingerprint travels inside Google's answer engine. It doesn't travel anywhere else.

When that vendor generates an answer using their own systems, the record transfers. Fine. That's just architecture.

But ChatGPT didn't build that record. Perplexity didn't build that record. Claude didn't build that record. Whatever profile *they* hold on your domain — assuming they hold one at all in any structured sense — was assembled from a completely different substrate. Common Crawl snapshots. Licensed data deals. Whatever their retrieval layer scrapes at query time. Reddit and Wikipedia weightings that nobody has documented. Bing's index in Copilot's case. Model training cutoffs that leave your last two years of work invisible until the next refresh.

The industry has quietly been talking about "AI visibility" as if it's one thing. It isn't. It's at least four different things, held by four different vendors, built from four different data pipelines, refreshed on four different cadences. And the confidence with which people are selling "GEO strategies" as if optimising for one optimises for all is not supported by anything I've seen in the actual research.

Your Google fingerprint travels inside Google's answer engine. It doesn't travel anywhere else.

The variance research nobody's citing yet

There's an arXiv paper that landed this week that I think is going to matter more than the industry realises. It's a variance-components decomposition of LLM brand recommendations — 12,933 responses across 20 brands, 8 languages, 3 models. The kind of methodologically boring, statistically rigorous work that gets ignored because it doesn't have a headline in it.

The finding worth stealing: query language accounts for 26.5% of the variance in a single response. Brand identity accounts for 1.5%. The intraclass correlation for brand is 0.0146.

Read that number again. A single AI answer carries almost no brand-discriminating signal. The signal is buried under language variance, resampling noise, and brand-in-context interactions that together account for the overwhelming majority of what moves the output.

Which means the "fingerprint" as it exists inside answer engines outside Google is closer to statistical smoke than to a stable record. You can measure it. It moves when you're not looking. It moves differently in English than in German. It moves differently on GPT-5.2 than on Gemini 3 Flash. And no amount of repeat-sampling a single prompt narrows the confidence interval meaningfully — the researchers found that adding a sixth repeat past the fifth reduces relative-error variance by 0.0003.

This is the substrate the industry is selling dashboards against.

What Google admitted, and what it implies

The Forrester piece flags something Google said in a completely different context that reads differently once you're thinking about answer engines. Google's site reputation abuse guidance describes systems that evaluate whether a section of a site is independent or starkly different from the main content, and measure those sections independently — so a subsection can stop inheriting site-wide signals.

Forget the spam framing. Notice what it admits about the general architecture. Authority isn't held as one number for one domain. It's assessed at variable granularity, and the granularity can change without you knowing.

Now layer the answer-engine transition on top. If Google can already re-scope which parts of your domain inherit which signals inside its own ranking system, what happens when those signals feed a generated answer that surfaces one specific paragraph from one specific page? The unit of inheritance isn't the domain any more. It's the passage. And the passage doesn't inherit domain-level trust in any stable way you can watch.

This is why the John Mueller comments this week about "crawled — currently not indexed" being partly a quality signal, including for low-quality AI-generated content, matter more than they look. Google is doing quality assessment at page level, holding some pages out of the index entirely, and those held-out pages don't feed AI Overviews either. Your domain fingerprint might be pristine. Your specific page can still be invisible.

The fingerprint is real. Its portability is the fantasy.

Here's where I'd push back on the way this conversation is going in the industry more broadly.

The fingerprint is real. Forrester's right about that. Everything on his list exists, is measurable, and has been the substrate of professional SEO for a decade and a half. I've spent 18 years shaping records exactly like the one he describes for clients, and I can tell you the record is granular and it does drive output.

The fantasy is that the record is a portable asset. That a strong Google fingerprint carries into ChatGPT visibility. That E-E-A-T signals mapped to Search Quality Rater Guidelines mean anything to Perplexity's retrieval layer. That schema coverage optimised for rich results does the same work when Claude is deciding whether to cite you.

Some of it transfers, some of the time, in ways nobody has cleanly measured. The IQRush research I wrote about a few weeks back suggested that most AI citation movement is statistical noise. This week's variance decomposition paper says the same thing with better methodology. The signal is there. It's just an order of magnitude weaker than the noise floor most tools are pricing as if it's clean data.

The honest answer is that we have one vendor's fingerprint carried into one vendor's answer engine, and everything else is guesswork with a dashboard on top.

Why this matters for how you spend money right now

If you're a UK business trying to work out where to put budget between now and Q1, the practical implications shake out roughly like this.

For Google's AI surfaces — AI Overviews and AI Mode — the fingerprint transfers. Traditional SEO fundamentals are load-bearing. Link authority, technical hygiene, Core Web Vitals, schema, credentialed authorship. The correlation between conventional ranking signals and AI citation inside Google is strong enough that most of the "GEO tactics" being sold to you are just SEO tactics with a new label. I've said this before and the research keeps confirming it.

For ChatGPT, Perplexity, Claude, and the rest, the fingerprint you built at Google doesn't cleanly transfer. What matters instead is a mix of: whether you're in the training data at all, whether you're cited in high-authority third-party sources those models weight heavily (Wikipedia, Reddit, major publishers), and whether their retrieval layers surface you for the specific query. None of those levers are the ones a traditional SEO agency knows how to pull. Two of them look a lot more like PR and brand-building than technical optimisation.

For measurement — and this is the bit almost nobody wants to say plainly — you cannot reliably measure your position in most of these systems yet. The variance is too high. The vendors publish nothing useful. The dashboards charging you four figures a month for "AI visibility monitoring" are, in a lot of cases, showing you noise with confidence intervals so wide they're closer to horoscopes than analytics.

None of this makes the work pointless. It makes the pricing of confidence wrong.

The honest limits

I'm making a couple of claims here that could turn out to be too strong. Worth naming.

Claim one: the domain fingerprint is Google-specific and doesn't transfer to non-Google answer engines. This is *mostly* true but not cleanly true. Bing's index feeds Copilot. Perplexity does some live retrieval that would pick up traditional web signals. There are indirect channels where E-E-A-T-shaped work eventually reaches other systems through third-party sources that got the signal from Google-shaped web. So the fingerprint isn't sealed inside Google. It's just that the transfer is lossy, indirect, and delayed in ways that don't respond quickly to your optimisation work.

Claim two: the variance research means AI visibility measurement is broken. The paper I'm leaning on is one study on one corpus with three models and eight languages. It's rigorous but it's not the final word. It could turn out that with better sampling design — more paraphrases, more languages, fewer repeats — you can extract stable brand-level signal from these systems. The researchers themselves suggest reliability rises to about 0.36 at the full crossed design. That's still low, but it's not zero.

Where I'd genuinely concede: for very high-volume, very well-established brands, the signal-to-noise ratio probably looks better than it does for the mid-market UK businesses I mostly work with. If you're Nike, "am I cited in AI answers about running shoes" is probably measurable. If you're a mid-sized B2B services firm in the Midlands, it isn't yet, and pretending otherwise is how consultants lose credibility.

What this leaves you with

The fingerprint is real. Its inheritance inside Google's answer engine is architectural and reliable. Its portability across every other answer engine is a marketing claim, not an engineering reality, and the research is starting to say so with numbers.

That means the honest allocation of budget and attention for the next two quarters looks less like a "GEO transformation programme" and more like: keep doing the SEO fundamentals that build the Google-shaped record properly, invest in the third-party citation surfaces (Wikipedia, high-authority publications, Reddit where appropriate) that non-Google systems weight independently, and stop paying premium prices for measurement tools that are showing you variance dressed as signal.

Forrester ended his piece by saying the honest answer to the fingerprint-inheritance question changes depending on which system you mean, and in one case nobody outside the lab actually knows.

I'd go one step further. In most cases, nobody inside the lab knows either. The systems weren't designed to be legible, and the industry building tools to make them legible is running about two years ahead of what the underlying data can support.

That's the loop we're in. Keep the fundamentals honest, and price the rest of it accordingly.