TL;DR
A premium needs three things: the good must be scarce, the buyer must be able to tell it from the substitute, and the buyer must actually be short of it. First-hand experience passes the first test and fails the other two.
Experience is scarce in the doing and abundant in the telling — and the telling is the only part a machine ever sees. The first-person markers the 2026 playbook tells you to add are the cheapest output a language model produces.
Answer engines are not short of experience. They are short of settlement: a claim they can commit to without trusting the person who made it. Exactly one class of experiential output supplies that, and it is not the class the playbook maximises.
The premium the industry has already priced
The argument arrived fully formed and is now close to consensus. Graphite’s Q1 2026 sample of 55,400 English-language articles from Common Crawl put primarily AI-generated articles at 49.9% of everything published, a figure hovering near half since early 2025. If half the web is machine-made, the reasoning goes, whatever a machine cannot produce becomes the scarce input, and whatever is scarce commands a premium. First-hand experience is what a machine cannot produce. Therefore: publish it, and collect.
The supporting evidence is real and recent. Google added Experience to E-A-T in December 2022 and has leaned on it since. On 6 May 2026, Google VP Hema Budaraju announced that AI Overviews would surface first-hand perspectives from forums, social platforms and public discussions in a dedicated block labelled Expert Advice or Community Perspectives, each entry carrying the creator’s name, handle or community. That is a search engine building furniture to display personal experience. Anyone arguing that experience does not matter is arguing with the interface.
So the advice writes itself, and by mid-2026 every serious guide carries the same list: original photography, first-person statements with dates and quantities, named authors, documented outcomes from real deployments. Most of it improves a page. But underneath sits one load-bearing assumption, stated in a sentence and never examined. A widely circulated 2026 guide to AI Overview citations puts it plainly: first-hand experience contains specific details that generic content lacks, and AI systems can detect these markers.
That sentence is the entire strategy. If it is true, the premium is collectable by writing differently. If it is false, several thousand businesses are paying a premium price for a signal nobody reads. The test is not difficult, and the answer changes what you should spend a link building budget on for the next two years.
What a premium actually requires
A premium is a price above the going rate for the ordinary version of a thing. Scarcity alone does not produce one. Three conditions have to hold together, and each one fails in a different way.
Scarcity has to hold at the point of sale rather than in production: a thing that is expensive to make and free to copy is scarce only to the maker. The buyer has to be able to tell the premium version from the ordinary one — the condition hallmarks, warranties and inspections exist to satisfy, and the one that quietly does all the work. And the buyer has to be short of it, because a premium is paid for whatever is binding. A restaurant short of chefs does not pay extra for flour.
First-hand experience passes the first test in an interesting and misleading way, and fails the other two outright.
What is the authenticity premium?
The authenticity premium is the extra citation and ranking value that first-hand experience is expected to command now that machine-written text is abundant and cheap. The demand behind it is real; the field has misidentified what the demand is for.
Scarce in the doing, abundant in the telling
Somebody has to fit the boiler, run the trial, sit through the tribunal, break the machine. That is genuinely scarce. But scarcity has to survive the transfer to the reader, and what crosses is not the experience. What crosses is a description of it, and a description is text.
This is Arrow’s old problem with information as a good: you cannot judge its value until it is disclosed, and once disclosed it need not be bought. Publish what you learned and the finding inside the description becomes freely restatable by people who did none of the work. Worse, those restatements corroborate each other. The tenth person to repeat your finding is adding to the pile of independent-looking sources that agree, and none of them fitted a boiler.
For example, a heating contractor who publishes that a particular flow sensor fails first has, within a month, three competitors saying the same thing in their own words. The contractor bore the whole cost of discovering it and holds no part of the resulting consensus. The scarce input arrives at the engine wearing exactly the same clothes as the abundant one.
The markers are the cheapest thing a model makes
The second condition is where the playbook breaks. Consider what the guides instruct you to add: first-person voice, sensory specificity, a date, a unit count, an admitted mistake, a confident aside about what surprised you on site. Every item is a property of style. Style is the thing generative models were optimised on. Asking a language model for a paragraph that reads like lived experience is not a hard prompt; it is close to the easiest thing you can ask it for.
The evidence that people cannot tell the difference is not contested. An academic study of consumers asked to classify AI-generated fake reviews measured detection accuracy at 53.2% — a coin flip, and worse on negative reviews. Machines do not do obviously better at the thing that matters here. Graphite validated three separate detectors (Pangram, GPTZero and Copyleaks) for its 2026 study and reported false-positive and false-negative rates below 2%, while noting in the same document that there is considerable disagreement about whether reliable detection is possible at all.
But notice what even a perfect detector would establish. It would tell you whether a text was machine-written, and nothing about whether the events described happened. Only the second question is what anybody means by authenticity. No vendor has shipped a classifier for whether a person actually did the thing they wrote about, because the evidence for that is not in the text. It never was.
There is no experience detector. There is a style detector, and style is the one thing a model copies perfectly.
This is the ordinary structure of a market where quality is unobservable at the point of purchase. The resolution is never better quality — the honest seller who tries harder is invisible by construction. It is always an external check: a hallmark, a warranty, a third party with something to lose. Consumers worked this out first. BrightLocal’s 2026 local consumer survey found the strongest trust factor for a review is not its star rating (42%) but whether other reviews echo it, at 56%. They grade agreement.
The engine is not short of experience
AirOps analysed 548,534 pages retrieved by ChatGPT across 15,000 prompts and found only 15% of them cited: 85% of what the model read never appeared in the answer. Graphite’s companion study found 86% of articles ranking in Google Search and 82% of those cited by ChatGPT and Perplexity were human-written, despite machine-written articles making up around half of what is published. The flood is real and mostly not reaching the answer.
An engine assembling an answer is not rummaging for material. It is discarding material. It has more plausible, experience-flavoured candidates than it can use, and its problem is choosing between them — including between two accounts that disagree. That is the thing it cannot do. It has no instrument for adjudication, so when sources conflict the behaviour is to hedge, attribute the claim to nobody in particular, or drop it.
The web has never carried more first-hand accounts. What it has less of than ever is any way to settle which of them is right.
That is the binding constraint, and it tells you what the premium is for. It is not paid for experience. It is paid for claims that can be settled without trusting you — and only some of what experience produces qualifies.
The five claim classes
Experience is not one output. It produces five kinds of statement, and they behave so differently that treating them as one category is the root error. For each, only two questions matter: what would settle it, and who else could produce it.
| Claim class | What it asserts | What would settle it | Who else can produce it |
| Texture WORTH LEAST | How it looked, felt, sounded. The app felt sluggish by the third day. | Nothing. There is no external check for an impression. | Anyone, including a model, at no cost. |
| Judgement HEDGED OR DROPPED | What you would advise. The wrong tool for a team of twelve. | Nothing. It is a recommendation, not a fact. | Any competitor with an opinion. |
| Occurrence NEEDS A RECORD | That you did the thing. We tested forty units over six weeks. | A third-party record: register entry, invoice, programme, certificate. | Anyone can claim it. Nobody can check it from your page. |
| Measurement OWNERSHIP ENDS AT PUBLICATION | A number produced by a stated procedure. 62 dB at one metre. | A disclosed, repeatable method. | Any rival with the same equipment — then everyone, by restatement. |
| Incidence THE ONLY DURABLE CLASS | A rate across a population you operate. 11% of 340 installs failed at the flow sensor within 18 months. | Your population, plus a disclosed method and a date. | Nobody without your operations. |
Now read the standard advice back against the table. Original photography, first-person voice, sensory detail, a personal narrative a quality rater can follow: those are rows one and two. The 2026 authenticity playbook is a set of instructions for maximising the two classes of claim that nothing can settle and anyone can fabricate. The advice is not wrong about what humans enjoy reading. It optimises the part of your experience that has no evidentiary standing.
Why measurement stops being yours the day you publish
A measurement is the first genuinely citable class, and the one most technical publishers reach for. But a number produced by a repeatable procedure is reproducible by anyone who owns the procedure — and once published, reproduction is unnecessary. The figure detaches immediately and circulates on its own. What stays attached is the method, if you disclosed one, and the series, if you repeat it. The number is a flow; the method and the series are the stock.
Why incidence survives
A rate across a population cannot be reproduced without the population, and the population is your operations. A veterinary group with sixty-one practices can state a post-operative complication rate no journalist, rival or model can generate from any amount of reading. The same group’s page explaining that its vets have seen it all states nothing. One is evidence, the other atmosphere, and they cost about the same to publish.
What a signature can and cannot say about your Tuesday
There is an obvious response: provenance. C2PA Content Credentials — cryptographically signed records of how a file was created and edited — are now an ISO standard, backed by camera manufacturers and platforms. If the problem is that anyone can claim to have been there, sign the evidence that you were.
It works, within a narrower boundary than it first appears. A camera signature establishes that these bytes came from this device at this time and have not been altered since — a real upgrade for exactly one row of the table, moving an occurrence claim from unverifiable assertion to attested capture. It says nothing about what the photograph shows, whether the test behind it was competently run, or whether the caption is accurate.
Notice where the upgrade lands. Provenance attaches at the moment of capture — which means it is available for the class of claim that is worth least and unavailable for the class that is worth most. A fire door inspection produces a photograph you can sign. A failure rate across eleven thousand inspections produces a spreadsheet, and no signature on it tells a reader whether those inspections happened or were recorded honestly. The most valuable thing your experience generates is the thing provenance reaches last.
Key takeaway
Signing raises an occurrence claim from unverifiable to attested, and does not reach an aggregate at all. If the value of your evidence rests on what happened across a population rather than on which file it happened in, provenance is not the intervention — a disclosed method and an independent restatement are.
The other thing that is not scarce: bought originality
If experience is not the moat, the field has a second answer: original data. Run a study. This is now the dominant format — one 2026 compilation of digital PR practice put the share of specialists using data-led campaigns at 95.9%, and data-driven work at 39.6% of analysed campaign coverage, roughly double the next format. Cision’s media surveys have shown for years that original research and exclusives materially improve the odds of coverage, and around two-thirds of journalists prefer pitches backed by data.
But look at how most of it is produced. A panel survey of five hundred qualified B2B respondents costs somewhere between $1,500 and $5,000 and takes a fortnight. That is a purchase order. Your competitor can buy the same panel on the same topic next quarter, and will. Bought originality is priced, so it is available to everyone who can pay — the definition of not scarce. It still earns links, because it is still useful, and it belongs in the mix alongside reactive expert commentary and journalist request platforms. What it does not do is create a position anyone has to come to you for.
The middle route: records somebody else already keeps
There is a third source of population data, and in the UK it is badly underused: the state. On 11 March 2026, Sentry Fire Safety Group published A Burning Issue, built from Freedom of Information responses covering around 89% of England’s local authorities. It reported that 63% of flat entrance doors and 67% of communal doors failed the FD30 thirty-minute standard, and that only 46% of flat entrance doors had been inspected even once since mandatory annual checks began in January 2023. The findings were reviewed at an industry and policy meeting, drew comment from the chair of the All-Party Parliamentary Group on Fire Safety and Rescue, and were reprinted across the trade press.
The operations were somebody else’s; the records were public; the work was in the asking, the cleaning and the willingness to publish a number that made a sector look bad. Note what the report conceded: housing associations are exempt from FOI, so the picture was partial. Stating that limit is not a weakness. It is what makes a figure quotable rather than arguable, and why the technique travels into international link building wherever a comparable disclosure regime exists.
The Evidence Yield
Most businesses already run the operations that would produce an incidence claim, and almost none can produce one. The reason is arithmetic rather than ambition. Pick the single quantity you would most like to own — the number a journalist would ring you about — and calculate:
The Evidence Yield calculation
Evidence Yield = operations in the window that left a structured, retained record of that quantity ÷ operations performed in the window.
A structured record means a field with fixed values, retained and comparable. A photograph is not a record of a rate, a free-text note is not a field, and an engineer’s memory is not retained.
For example: 340 installs last year, flow-sensor outcome captured in a structured field on 41 of them. Yield = 12%. You do not have a claim. You have an anecdote with a large denominator you are not allowed to use.
Sample size then determines what the figure can be used for, and the thresholds are lower than most assume:
- Under 30: an anecdote. Publish it as a story, not a statistic, and do not put a percentage on it.
- 30 to 100: indicative. Publishable with the method and limits stated; a trade editor will run it framed as a survey of a named sample.
- 100 to 1,000: trade evidence. Journalists will use the figure with attribution; rivals will start quoting it back at you within a quarter.
- Over 1,000: a reference series, but only if you repeat it on a fixed date. A single large study is a news story; the same study run annually becomes the number the sector uses.
Now the part that hurts. Yield is a function of what you decided to record, not what you did. The denominator is your history; the numerator is your filing discipline. For most firms the first runs into the thousands and the second sits near zero, and no amount of better writing recovers the gap.
The retention rule
You cannot publish an incidence claim you did not decide to record twelve months ago. Every quarter you run without a structured field is a quarter of evidence you have permanently spent.
The corollary is a taxonomy freeze. If your failure categories or outcome labels change mid-year, your comparable sample resets to whatever came after the change.
What this changes about acquiring links
If the scarce good is a claim that settles, the job of an earned placement changes. A link has always done two things — pass authority and put your name somewhere useful. Against a settlement problem it does a third: an independent publication restating your figure is the external check a first-person account can never supply for itself. That is why what link building is fundamentally for has quietly shifted under the industry’s feet. You are not only buying reach. You are buying corroboration for a claim that would otherwise be a stranger’s assertion.
Screen outlets for numeric uptake
Authority score is the wrong first filter here. What you need to know is whether an outlet reprints figures. Pull the last ten data stories a target publication ran and check whether the number and its source survived into the copy, or dissolved into a phrase like research suggests. The second kind is pleasant and worth having; it settles nothing. A modest trade title that prints the figure, the sample size and the source is doing more for you than a national that mentions your brand and rounds the number away. This is a different screen from the ones used for guest posting, niche edits or sponsorship placements, and it is uncorrelated with all of them.
Buy the series, not the report
A figure published once is a news story with a decaying tail. The same figure, published on the same date every year with the same method, becomes the reference — and reference status earns links passively for as long as the series runs. It is the most under-priced decision in the exercise, because it costs nothing at first publication and cannot be bought retrospectively. A competitor can match your study. They cannot match your fourth annual edition until year four.
Prefer carriers that do not disappear
Trade bodies, regulators, professional institutes and insurers are the highest-value destinations for an incidence figure: their pages are rarely paywalled, rarely deleted, and their restatement converts your claim into the sector’s claim. Once a trade body quotes your failure rate, whether to trust you has been answered by somebody with no interest in your winning. Pitch them even when they do not link — worth weighing when you are measuring entity authority rather than counting referring domains.
Write the claim to travel. A figure carrying its denominator, its date and its method is restated intact; a bare percentage is restated as a rumour. This matters for the same reasons that shape what drives AI product recommendations and how lost AI citations are recovered: a claim carrying its own qualifications survives compression; one that does not gets hedged out of the answer.
Where this argument is weakest
The strongest objection is not theoretical. It is that the premium is being paid right now, at scale, to exactly the classes marked red.
Tinuiti’s Q1 2026 AI Citations Trends Report, produced with Profound across nine commercial categories and seven AI platforms, found social media’s share of AI citations topping 9% by January, with Reddit dominant and its share up at least 73% across tracked categories. Separate analyses put Reddit at 46.7% of Perplexity’s social citations, and Conductor found it increasingly cited as the only source, with sole-source citations up 31% since October 2025. And Google has now built a dedicated block for exactly this material. Anonymous people describing how something felt are winning citations in volume — texture and judgement, which by the table above should be worth nothing.
The observation is correct and it is the best evidence against this argument. Four things bound it.
First, the unit is the thread, not the post. What gets cited is a convergence of independent accounts carrying vote counts, account histories and moderation — a corroboration mechanism doing exactly the settlement work described above. The comment supplies raw material; the platform supplies the check. You cannot install that mechanism on your own domain, and a testimonials page is not a weaker version of it but a different object entirely.
Second, it is contracted supply, not open-web merit. Reddit’s data agreements with OpenAI and Google are reported at around $130m annually in aggregate, and the behaviour across engines is wildly inconsistent: Tinuiti’s January 2026 data shows Reddit above 5% of ChatGPT citations and 0.1% on Gemini, with the lower figure belonging to the company that pays for the data. Whatever drives that gap, it is not a measurement of how authentic the content is.
Third, it is unstable in a way a budget cannot absorb. Semrush’s three-month tracking caught ChatGPT’s Reddit citations collapsing from roughly 60% of prompt responses to around 10% in mid-September before recovering. A premium you cannot forecast one quarter to the next is not a premium. It is a lottery that happens to be paying out.
Fourth, and decisively for a business, the action does not change. Grant the objection in full and you still cannot become a community. The participation route is not yours to control, is removable at a moderator’s discretion, and is now squarely regulated: the FTC’s rule at 16 CFR Part 465 covers testimonials from people who do not exist, and the UK’s DMCC Act 2024 fake review provisions have been CMA-enforceable since April 2025 with penalties reaching 10% of turnover. Participate honestly where your customers are — but what you can own, price and repeat is a record of your own operations.
The objection kills the claim that texture is worthless. It does not touch the claim that texture is not yours.
A worked example: Thurlby Passive Fire
Thurlby Passive Fire is a Nottingham firm with £9.7m of turnover and 74 inspectors, carrying out around 14,000 fire door inspections a year for social landlords, two NHS trusts and a university estate. Over three years it published 46 articles to the standard specification: named authors, original site photography, first-person accounts of what inspectors find behind communal doors. The agency invoice came to £31,000. Across 120 sector prompts on four engines, the site was cited in three.
The diagnosis took an afternoon. Their Evidence Yield on the quantity that mattered — why a door fails — was 7%. The inspection app captured a pass/fail flag as a structured field and everything else as free text and photographs, and the failure-reason list had been revised twice in two years, so even the free text was not comparable. They had inspected roughly 42,000 doors and could not state a single rate about any of them.
Sixteen weeks
A fixed nine-field failure taxonomy replaced the free text, cut from a proposed fourteen after inspectors timed it: fourteen fields added about ninety seconds per door, nine added thirty-five. Three months were backfilled by re-coding photographs. A legal and client-consent pass ran in parallel. It published as a dated annual with the full method and the raw counts attached.
The findings covered 11,908 inspections across 214 buildings over nine months: 61% of doors failed at least one FD30 criterion; damaged or missing intumescent strips were the largest single cause at 31%; and 8% failed on a defect introduced by previous remedial work. That last figure was the one nobody else could produce, because it required the same firm to have inspected the same door twice — structurally unavailable to any surveyor, journalist or model without a repeat-visit population.
What it returned, and what it cost
Twenty-seven referring domains in eleven weeks, including a trade body, an insurer’s risk bulletin and two trade titles that printed the figure with the sample size intact. On the 120-prompt panel, citations went from 3 to 31, concentrated entirely on failure-cause questions.
The honest ledger is less tidy. Commercial prompts did not move: nobody asking what a fire door inspection costs in Nottingham was any likelier to see them, because the data answered a question buyers were not asking. Enquiries rose 9% and the firm cannot attribute that to anything. Legal review took six weeks and blocked two of five findings; one landlord objected and the breakdown by landlord type was dropped.
A competitor restated the 31% figure within five weeks without a link, and by month four two consultancies were quoting it as an industry number with no attribution. That is the ownership decay the table predicts for a published measurement, and it is not a failure of the work — the number stopped being theirs the day it became useful. What stayed theirs was the method, the repeat-visit population and the date in the calendar. The recurring cost is permanent: thirty-five seconds a door across 14,000 doors is about 136 inspector-hours a year, and the first report consumed £19,000 of analyst time, more than the three years of articles it replaced.
What to do on Monday
- Name the single quantity you would most like to own — the number a journalist would ring you for. One, not five.
- Count the denominator: how many operations touched that quantity in the last twelve months?
- Count the numerator: how many left a structured, retained record — a field with fixed values, not a note or a photograph?
- Divide. If the Evidence Yield is under 20%, stop commissioning articles about your experience this quarter and fix the capture instead.
- Freeze the taxonomy in writing and refuse to revise it for four quarters. A mid-year change resets your comparable sample.
- Run your ten target publications through the numeric-uptake test: in their last ten data stories, did the figure and its source reach the copy?
- Put next year’s publication date in the calendar before you publish this year’s. The series is the asset; the report is the announcement.
- Clear consent, confidentiality and anonymisation before you build the dataset, not after — and expect to lose one finding to it.
Two notes if you are running this alongside an existing programme. An incidence figure reads as evidence to a regulator as well as to a journalist, which is unusually safe ground next to tactics that later need manual action recovery or a negative SEO defence later. And if you publish in a European market, check your disclosure obligations under the EU AI Act content rules before pairing generated illustrations with real operational data.
The premium is a filing decision
The industry has spent two years telling businesses to sound more like people who have done things. Engines were never listening for it, and the tools that track what actually gets cited will keep reporting the same disappointing number for anyone who takes that advice literally. An answer engine needs a sentence that survives being checked. It does not care whether the sentence is warm.
The uncomfortable part is the lead time. A page can be rewritten on a Tuesday; a rate cannot. The claim you want next spring had to start being recorded last autumn, which makes this a decision for whoever owns the operational systems rather than the content calendar — an unusual reporting line for link acquisition, and the correct one. It also means the 2026 link building statistics that matter most to your own programme are the ones you have not started collecting yet, whether you are chasing featured snippet placement, local citation coverage or visibility inside the new AI browsers.
Authenticity, priced honestly, is not a voice. It is the decision to keep the record while the thing is still happening.
