TL;DR
It is true that a model cannot manufacture first-hand experience. It is not true that it struggles to write experiential prose. It writes that better than most people, because experiential prose is high-probability text and a model is a likelihood machine.
What a model cannot generate is a particular that would be wrong if it were not true. Fluency is the ability to produce the likely; experience is the licence to assert the unlikely.
The signal also has to survive extraction. Answer engines lift sentences, not pages, and paraphrase removes voice while preserving numbers. In one study of 3,200 cited passages, plain narrative prose was reproduced word-for-word 12% of the time and statistic lines 61%.
So grade your sentences by exposure, not by specificity: who could prove this wrong, and with what? The top two rungs of that ladder cannot be built on your own domain, which is where the link programme comes in.
1. The claim is right. Almost everything built on it is wrong
The consensus position going into 2027 is easy to state, and you can find it in almost identical words across the guidance published this year. Search and answer engines are drowning in generated text. The one input that cannot be synthesised is having actually done the thing. Therefore: add first-hand experience markers — original data, direct experimentation, first-person case studies, primary-source interviews — and you hold the one position a competitor with a content pipeline cannot take from you.
The first sentence of that argument is correct. A model has no hands, has been nowhere, and has measured nothing. The conclusion drawn from it is where the whole thing comes apart, because it quietly assumes that experience and the description of experience are the same asset. They are not, and only one of them is scarce.
What is a first-hand experience signal?
A first-hand experience signal is anything on a page that gives a reader — human or machine — grounds to conclude the writer was present for what they are describing. The important word is grounds. Experience is not something you can transmit directly; it is a conclusion someone else draws. What you can transmit is evidence that makes that conclusion cheap, or prose that merely asserts it.
Consider what a language model is optimised to do. It predicts the next token, and its training and evaluation regimes reward confident, plausible continuations. OpenAI’s own September 2025 paper on why models hallucinate makes the point bluntly: benchmarks that penalise I do not know teach models that guessing pays. A system built to produce the plausible is extremely good at first-person narration, because first-person narration is one of the most heavily represented registers in its training data.
Ask a model to write four hundred words about the third time a bearing failed on a night shift and it will do it fluently, in the trade’s own idiom, with the right amount of weary understatement. Ask it for the torque figure at which that bearing actually shears, on a component nobody has published data about, and it will either decline or invent one. That is the real boundary, and it does not run between human and machine writing. It runs between the assertable and the checkable.
Key takeaway. Fluency is the ability to produce the likely. First-hand experience is the licence to assert the unlikely. The discipline has spent three years teaching people to write the likely, in the first person.
2. Nobody is reading your page
Before asking what an experience signal should contain, it is worth asking who receives it. The playbook assumes a reader who takes in a page as a whole and forms an impression of the person behind it. Neither of the two systems that matter works that way.
The ranking layer has no experience input
E-E-A-T is not a ranking factor. There is no experience score, no field in an index where it is stored, and no rater who can move a page. The Search Quality Rater Guidelines — 182 pages, last revised 11 September 2025 and still the current version — describe the target the ranking systems are tuned towards. They are an evaluation instrument, not a lever.
More usefully, the guidelines say something about experience claims specifically that the playbook never quotes. Raters are told to base their assessment on the main content itself, on reputation research and on verifiable credentials, rather than on what the site says about itself. And where claims of personal experience or expertise appear overstated or included just to impress visitors, the instruction is to apply a Low rating. Search Engine Land’s reporting on the revision framed it as a crackdown on exaggerated experience claims, which is precisely what it is.
Read that carefully. The central move of the E-E-A-T experience playbook — assert that you have done the thing, prominently, in the first person — is the exact move the guidelines instruct raters to discount, and in its more enthusiastic forms, to route towards the sharper end of quality enforcement. Everything you have read about technical implementation of author signals is downstream of a document that tells its readers not to take those signals at face value.
The answer layer reads sentences, not pages
The second system is stranger. Retrieval-augmented answering splits into three separate decisions that get collapsed in most advice. Chunking cuts your page into passages at index time. Retrieval decides which passages surface for a question. Passage extraction picks the individual sentence inside a surfaced passage that goes into the answer. Only the last of those is about your writing, and it operates on a unit no editor has ever worked in.
The measurements bear this out, and the published citation data is consistent on the point. Ahrefs, looking at 174,048 pages across 560,346 AI Overviews in December 2025, found a Spearman correlation of 0.04 between word count and citation position — effectively nothing. More than half of cited pages ran under a thousand words. By March 2026 the same firm found that top-ten organic results accounted for only 38% of AI Overview citations, down from 76% eight months earlier, so the page-level judgement that used to route everything now routes considerably less. Whatever still connects rankings and citations, it is not a channel through which an impression of a whole document travels.
Key takeaway. An experiential page is a category error twice over. The ranking layer has no input for it and the answer layer never sees a page. Experience is carried by a sentence or it is not carried.
3. Paraphrase destroys voice and preserves number
This is the mechanical heart of it, and it is measurable. When an engine lifts a passage, it either reproduces it or rewrites it — and which of those happens depends almost entirely on what shape the passage is in.
MaxAEO’s study of 3,200 cited passages sorted pulls by format. Plain narrative paragraphs — flowing prose with no number, definition, list or table structure — took just 4% of pulls, and when they were used, only 12% survived word-for-word; the rest were reworded past recognition. Statistic lines came through verbatim 61% of the time and table rows 58%. Across the whole sample around 45% of quoted passages were reproduced with no more than trivial edits, but that average conceals the split entirely. The median quoted passage ran 28 words, with 70% of verbatim pulls falling between 15 and 45.
Now think about what a rewrite does to a sentence. Take we found the bond failing on slabs that had passed the test and paraphrase it. The first thing to go is we found, because it is the only part that can be deleted without losing information. What remains is the bond, the slabs and the test. Voice is what a paraphrase removes at no cost. A number is what it cannot remove at all.
Which means that every hour spent making a page sound more experiential is spent on the one property of the page guaranteed not to travel.
Hedging is the second casualty
There is a related finding worth sitting with. Engines quote a source when it is concise and definite, and paraphrase it when it is hedged or wordy. Experiential writing is structurally hedged — that is what makes it honest. In our experience, we have generally found, it tends to. Each of those is a truthful signal of the limits of a sample, and each is an instruction to the extraction step to rewrite rather than quote.
The resolution is not to strip the honesty out. It is to put the hedge somewhere other than inside the sentence that carries the claim. State the finding flatly, then state its limits in the following sentence. The claim becomes quotable; the caveat stays on the page for the reader who has arrived to check.
Does writing in the first person help or hurt AI citation?
Neither, directly — but it competes for the same words. First-person framing is removed by paraphrase, so a sentence built around it arrives in an answer stripped of everything except whatever concrete content was left over. Where a first-person sentence contains no checkable particular, nothing survives at all. Write the particular first and let the person be the reason it exists.
There is a sting in this that explains a lot of disappointed reporting. When experience signalling works, it stops looking like experience. It arrives in someone else’s answer as a flat fact with your name attached to a number, which nobody on the team recognises as the thing they commissioned. Teams running a presence audit across engines often conclude the experience programme produced nothing, when what actually happened is that it produced exactly one sentence and that sentence is now in circulation without its adjectives, including inside the longer research reports engines assemble.
4. The Commitment Grade
So what survives? Not specificity as such — a model overproduces specifics. What survives is exposure: the property of a sentence that somebody outside your organisation could prove it wrong, and would have the means and the standing to do so.
Exposure is gradeable. Take any sentence from a page that claims first-hand knowledge and ask two questions in order: does this have a truth value at all, and if so, who holds the record that could contradict it? Five answers are possible, and they form a ladder.
| Grade | What the sentence does | Who could contradict it | What an engine can do with it |
| L0 — Assertion | States a quality rather than a fact. We have decades of hands-on experience. There is no proposition, so no truth value. | Nobody. There is nothing here to be wrong about. | Nothing. It can generate this sentence itself, in your voice, at no cost. |
| L1 — Private | States a real proposition, but the only record of it is yours. We have completed over 400 of these. | Only you. No outside party can check it either way. | Attribute it or drop it. In practice most engines drop it. |
| L2 — Joinable | Ties the claim to a public artifact — a standard, a spec, a published figure — so half of it is already checkable. | Anyone holding the artifact can check the half you did not produce. | Place your half against the public half and reproduce the pair. |
| L3 — Exposed | Makes a claim a named party holds records against, and would recognise as being about them. | Manufacturers, test houses, insurers, regulators, registers — parties who would know, and would say so. | Carry it as a live claim with an owner. Naming is how a model hedges what it will not assert itself. |
| L4 — Checked | The claim has already been confirmed or contradicted in public, in writing, by someone who is not you. | Already happened, on the record, under a name. | Treat it as settled enough to state, and cite whoever did the checking. |
The distribution is the finding. Practically the entire genre publishes at L0 and L1. The author bio, the years-in-the-trade line, the process narration, the case study with the client anonymised out of existence — all L0 or L1, all uncheckable, all cheap to produce without having done anything. L2 is the floor for carriage, because it is the first rung where a sentence connects to something the rest of the world can already see. It is a different axis from whether the content itself is verified — a signed page of L0 assertions is still a page of L0 assertions.
And the top two rungs cannot be built on your own website. L3 requires a party who holds records about your subject and is not you. L4 requires that party to have acted. No amount of on-page work produces either, which is the first practical consequence of the whole argument and the reason this is a link building problem before it is a content one.
What makes an experience claim citable?
Exposure, not detail. A citable experience claim is one where a party who is not the publisher could check it, would have the means to check it, and would be recognised as having standing if they disputed it. Detail without exposure reads identically to fabricated detail, because on the page it is identical.
5. The Commitment Ratio
The grade applies to a sentence. To apply it to an estate you need one number, and it is a ratio rather than a count, because the failure mode is not too little experiential writing. It is too much of it, holding up nothing.
The Commitment Ratio
1. Take your twenty most experience-heavy pages and split them into sentences. This is a text-processing job of about an hour, not a research project.
2. Count experiential framings. In our experience, we have found, having done this since, on site we see, from what we have seen. Call it F.
3. Count committed particulars: sentences at Level 2 or above — a specific claim joined to something outside the page that a party who is not you could check. Call it C.
4. The Commitment Ratio is C divided by F. Report both raw counts per thousand words as well, because a ratio of 1.0 built from one framing and one particular is not a programme.
5. Then measure the other end. Run forty buyer questions across five engines every month and count how many of the sentences reproduced about your subject are yours. That is carriage. Track what carries into follow-up turns separately.
Bands. Below 0.1, the experience on your site is decorative. Between 0.1 and 0.5, mixed. Above 0.5, committed — and rare enough that most estates have never measured it.
Run this on a page built to the standard playbook and the result is close to zero, usually by an order of magnitude. It is what happens when a genre optimises the assertion of evidence rather than the production of it. In our experience is a claim about evidence, not evidence, and it is one of the most reproducible strings in the entire discipline.
Why the ratio moves before the traffic does
The ratio is a leading indicator with a long lag behind it. A committed particular has to be published, crawled, chunked and then selected against whatever else answers the same sub-question, and the selection step is contested on grounds unrelated to quality. Expect the ratio to move in a week and carriage to move in a quarter. Teams who model the traffic side of this honestly already know that the rates in their reporting have stopped meaning what they meant; the counts are what remain clean, and the Commitment Ratio is a count of your own prose.
6. The strongest objection: fabricated specificity is cheap
The hardest counter to everything above is not that models cannot produce specifics. It is that they overproduce them, confidently, and that a fabricated particular sits on a page indistinguishable from a measured one.
This is not a hypothetical. Across 69,557 citation instances checked against CrossRef, OpenAlex and Semantic Scholar, hallucination rates ran from 11.4% to 56.8% depending on model, domain and prompt framing. Xu and colleagues, benchmarking thirteen models on citation generation in 2026, found rates between 14% and 95%. Sakai and colleagues identified close to three hundred published papers carrying at least one fabricated citation. The failure is not rare and it is not confined to weak models; the mechanism, as OpenAI described it, is that the training objective rewards a plausible continuation over an honest refusal.
Concede the point completely. A competitor with no data can publish a page of precise-looking figures this afternoon, and nothing about the page will give them away. The Commitment Grade does not fix that, because the grade was never scoring specificity. It scores exposure, and exposure is not a property of the page. It is a property of what the page connects to. Four things follow.
- The join is the test, and it fails compoundly. Ansari’s 2026 analysis of 100 hallucinated citations across 53 NeurIPS papers found 66% were total fabrications and 100% showed compound failure — several fields wrong simultaneously. Invention does not go wrong in one place. It goes wrong everywhere it touches something real.
- Fabrications do not replicate. Hallucination is stochastic rather than systematic: a model fills the same gap differently on each attempt. Measurement converges across independent attempts; invention diverges. That divergence is the same property that makes machine-generated patterns detectable at scale, and it is what a technical department finds when it tries to reproduce your figure.
- The asymmetry is in maintenance, not publication. Publishing one fabricated particular is free. Publishing the second one, which has to agree with the first, and the twelfth, which has to agree with both, is not. A series is what gets checked, which is why the cadence and consistency of what you put out carries more information than any individual page.
- The consequence is commercial, not algorithmic. A contradicted particular published under a trading name is not a ranking event. It is a customer with a printout, an insurer with a question, or a specifier removing you from a list. That is the actual restraint, and it does not apply to anyone operating without a name.
Here is the honest bound on the concession. All four of those are claims about the second year, not the first. In the first six months, fabricated specificity works, and anyone telling you otherwise is describing a market they would like rather than the one that exists. What the exposure ladder buys is not immunity from that competitor. It is a position that survives them.
7. The second objection: not all experience reduces to a number
The second real objection comes from everyone whose subject is not instrumented. A travel writer, a chef, a midwife, a furniture maker — their first-hand knowledge is genuine, valuable and largely qualitative. Told to publish committed particulars, they will reasonably ask which ones.
The objection is right about the arithmetic and wrong about the rule. The carriable unit is not a number. It is an indexical: a statement that points at a specific thing, in a specific place, at a specific time, that somebody else could go and look at. The second-floor rooms back onto the extraction fan is an indexical. It has a truth value, a party who could contradict it, and it survives paraphrase intact because there is nothing decorative in it to remove.
What does not survive is evaluation. Charming, worth the money, a real find — these are L0 in any sector, and they are what a model generates most easily, because judgements are the highest-frequency, lowest-risk output available to it. So the rule generalises cleanly even where the ratio does not: carry the indexical, drop the evaluative.
Key takeaway. A number is one kind of indexical, not the category. Anything that points at a particular thing another person could go and check does the same work — and evaluation, however sincerely felt, does none of it.
8. What this looked like in practice
Petherwick Resin Systems, a Nuneaton contractor laying industrial resin floors, turns over £9.4m with 34 installers. Their market is warehouse, food-production and pharmaceutical floors, and their commercial problem is bond failure: a floor that lifts at eighteen months is a six-figure remediation and a lost framework.
In October 2025 they ran the standard experience programme. Forty articles over five months, each bylined by a site supervisor, each written in the first person, each with process photography — £29,000 including the photographer. By February 2026 a forty-prompt panel across five engines named them twice.
The audit
The Commitment Ratio on those forty pages was 0.06. Across roughly 86,000 words there were 1,190 experiential framings and eleven committed particulars, nine of which sat at L1 — figures only Petherwick held. The pages read as though they had been written by people who knew the trade, because they had been. They contained almost nothing anyone could disagree with.
The audit also found the asset. British Standard 8204-6 covers synthetic resin flooring and, in the absence of a manufacturer’s own figure, sets the substrate at no more than 75% relative humidity measured by the hygrometer method in BS 8203. Petherwick’s supervisors had believed for years that the threshold behaved differently in cold weather. Their insurer had required a pre-pour record since 2019: hygrometer readings paired with a slab thermocouple, across 4,900 pours, with 61 recorded bond failures.
The publication
Between March and June 2026 they published one thing: the substrate window, a failure surface plotted against relative humidity and slab temperature. The headline claim was that a reading taken with the slab below 12°C under-reads equilibrium humidity by three to six points, and that 44 of their 61 failures came from slabs that had passed the test cold and would have failed it warm.
Every sentence in it sat at L2 or above, because every one of them joined to a published threshold that the Concrete Society itself has said some in the industry now regard as unrealistically low. The claim did not contradict the standard. It added a condition the standard does not state, which is the most contradictable thing a contractor can say.
By August 2026 the panel named them in 17 of 40 prompts. The detail that mattered: in 14 of those 17, the reproduced sentence was the flat form — the temperature figure with the firm attached as the source — and not one of the reproduced sentences came from the forty experiential articles. A resin manufacturer’s technical bulletin restated the window, and two trade bodies referenced it; the bulletin was the most valuable placement by a distance, because its readers hold thermocouples.
Four things that went wrong
- Volume fell twelve-fold and traffic went with it. Three committed pieces replaced forty experiential ones. The forty had been ranking perfectly well for long-tail queries; they simply were not being carried. For two quarters the firm was worse off on the number its managing director actually looked at, and the framework has no answer to that beyond naming the trade honestly in advance.
- Publishing a failure surface means publishing failures. Two main contractors asked whether 61 in 4,900 was industry-normal — a question with no public answer, because nobody else publishes. Exposure to contradiction is exposure. It does not arrive only from strangers.
- A competitor cleared L2 with invented figures and it worked for five months. They published a rival window with no dataset behind it. Nothing on the page distinguished it, and it was carried alongside Petherwick’s until a manufacturer’s technical team failed to reproduce it in August. Five months is a real result, and it is the strongest evidence for the objection above.
- Grade did not transfer across question types. The window was L3 and carried. The firm’s genuinely first-hand procedural knowledge — how to phase a pour around a live warehouse — never rose above L1 however it was written, because a scheduling judgement joins to no external record. Some real experience has no exposable form, and rewriting does not create one.
9. What it changes about who you buy from
The practical consequence of the ladder is that its top half is off-domain by construction, which turns an E-E-A-T question into a prospecting one. Three changes follow, and none of them is about domain authority.
Screen by who could prove you wrong
The conventional question about a placement is who reads it. The exposure question is narrower and much more useful: who reads it that could contradict me? A trade publication whose readers operate the same equipment is not an audience, it is a test — and a claim that survives it has been graded by the market rather than by you. A general-interest outlet with ten times the traffic exposes the same claim to nobody who could check it, and therefore carries no commitment at all.
This screen is close to orthogonal to everything in a standard prospecting stack. It does not correlate with authority scores, and you will not find it in a competitor backlink analysis, because it is a property of a publication’s readership rather than of its link graph. It has to be assessed by hand, which is exactly why it is still available.
Buy the parties who hold records
L3 needs a counterparty who observes something and would recognise your claim as being about them. In most sectors that list is short and unglamorous: manufacturers’ technical departments, test houses, insurers, professional registers, standards committees, procurement functions, and the trade press that still runs a corrections policy. These score badly on every metric in the usual tool stack, and they hold nearly all the exposure available in a given market. They are also slow, which makes them a poor fit for news-pegged outreach and an excellent fit for anything you will still be citing in two years.
What you are buying from them is not endorsement. It is the existence of a party who would be in a position to disagree — which is what converts your sentence from something you said into something that has been left standing.
Stop buying the things that cannot carry
The corollary is a short list of placements to stop paying for on experience grounds specifically. Author bios on other people’s sites are L0 wherever they sit. Expert-contributor columns with no particular in them are L0 with a masthead. Roundup quotes assembled from journalist request platforms are usually L0 or L1, because the format asks for an opinion rather than a record; they are worth buying for reach and for the link itself, which is a separate and legitimate purchase, but they will not move your Commitment Ratio by a point.
The same discipline applies to your own assets. An interactive calculator earns links by being useful, but it only earns carriage if the coefficients inside it came from somewhere and you say where. A data-led narrative presentation carries whatever its underlying figures carry and nothing more; the presentation layer, like the first person, is removed on the way out. Reviewing your wider tactical mix against the ladder tends to sort it into two piles very quickly — things that buy distribution, and things that buy exposure. Both are worth money. They are not substitutes, and the second is currently far cheaper than it should be.
10. What to do on Monday
- Run the Commitment Ratio on twenty pages. Split into sentences, count experiential framings, count L2-or-above particulars, divide. An hour of work. Expect a number below 0.1 and do not treat that as a failure — it is the base rate.
- Grade your ten strongest claims. For each, write down who could contradict it and with what. Any claim where the answer is nobody is not an experience signal, whatever it cost to write.
- Find the record you already keep. Somewhere in the business is a dataset generated as a by-product of operating: a compliance log, an insurer requirement, a warranty file, a service history. It exists because something else forced it into existence, which is why nobody has thought of it as content.
- Rewrite one page so the claim leads and the hedge follows. Flat finding in one sentence; limits in the next. Change nothing about the honesty, only the order.
- Name your entities in every sentence you want quoted. A passage that says it or this loses its subject the moment it is lifted out of the page. Replace pronouns with the actual noun in any sentence that carries a claim.
- Build a ten-name exposure list. Who in your market holds records against which your claims could be checked? Approach them with a finding, not a pitch — the ask is that they check it, and the placement follows from that.
- Set the panel running. Forty questions, five engines, monthly, recording which sentences come back and whose they are. Carriage is the outcome; everything else in this article is a leading indicator of it.
The consensus is right that first-hand experience is the input a competitor cannot synthesise. It is wrong about where that input lives. It does not live in the telling, which is the cheapest thing on the modern web, and it does not live in a page-level impression that neither system is built to form. It lives in the handful of sentences you are entitled to write because of something you did and someone else could check — and in the parties, outside your domain, who are in a position to check them.
