The Provenance Report: Why AI Discovery Content Fails Before It Is Written cover image

White paper

The Provenance Report: Why AI Discovery Content Fails Before It Is Written

AI assistants cite brand-owned pages more than any other source. Most brands write those pages from inputs every competitor shares. This report examines where discovery content comes from, and what it costs when the source is public.

By Anthony Gale — Co-Founder, Geoffy

Research

Executive summary

There is a widely held assumption about AI product discovery that the largest available dataset contradicts. The assumption is that AI assistants prefer community sources — Reddit threads, review platforms, forums — over brand-owned pages, and that brands are therefore fighting for scraps in their own category.

That is not what the data shows. Roughly 57% of AI citations go to company-operated web properties. Brand-owned content is the majority of what gets cited.

This should be encouraging, and for most brands it is not, because of what follows from it. If brand pages are the primary citation surface, then the competition for a citation is a competition between brand pages. And the inputs most brands use to write those pages are inputs every one of their competitors also has.

Four sources sit behind most ecommerce content briefs: the existing product page, Search Console, competitor pages, and general category knowledge. All four are public. Any competitor can reach them. So can any AI writing tool pointed at the category. The result is convergent content: pages that differ in wording and agree in substance. An assistant choosing between five convergent pages has no reason to prefer yours.

The material that cannot be copied is not public. It sits in four first-party records that most brands already hold and almost none use for content: support tickets, product reviews, returns reasons, and on-site search queries. Every one of them is readable through a documented API today.

This report sets out what AI assistants cite, why public inputs produce interchangeable pages, where the non-public material lives, and a four-question audit that can be applied to a content brief before anything is written. It also states plainly what the evidence does not support, because several claims circulating in this space are wrong.


1. What AI assistants actually cite

Three findings frame the argument.

Brand sites are the majority of citations. Profound analysed 11.84 billion citations between 16 April and 16 July 2026, across eight models (ChatGPT, Claude, Google AI Mode, Google AI Overviews, Gemini, Grok, Microsoft Copilot and Perplexity), classifying 3.02 million domains covering 98.3% of global citation volume. More than one citation in two goes to a brand site, with the overall figure at approximately 57%.

Social and UGC platforms take a much smaller and highly variable share. Google AI Overviews cites social sources at 15.3% and AI Mode at 14.4%, while Microsoft Copilot uses a social source roughly once in every 29 citations.

One caveat matters when reading that figure: “brand site” means any company-operated property, not necessarily the brand being asked about. A citation to a retailer or a manufacturer both count. And Profound has not disclosed how its prompt set was selected.

Reddit is the most-cited single domain, and that means less than it sounds. Profound’s earlier study with Reddit, published 10 November 2025 across four billion citations, found Reddit to be the single most-cited domain in AI answers at 3.11% share. It ranks first on a very long tail. Rank one is not dominance, and on Copilot the same domain sits around 31st.

Ranking no longer determines citation. Ahrefs examined 863,000 keyword SERPs and four million AI Overview URLs, published 2 March 2026. Only 38% of AI Overview citations came from pages ranking in the top ten, down from 76% a year earlier.

Put the three together and the picture is clear enough. The citation surface is mostly brand-owned. Getting there is decreasingly a function of where you rank. Which leaves the content itself doing the work, and raises the question of where that content comes from.


2. The four public inputs, and why each one fails

The failure here is structural rather than qualitative. These inputs do not produce bad pages. They produce the same pages.

The existing product page. The most common starting point for a rewrite, and the most obviously circular. Your product page is public, indexed, and already read by every competitor in your category. A page rewritten from itself adds phrasing, not substance.

Search Console. Records queries that produced an impression or click in Google Search. It cannot record a prompt typed into an assistant, because that string never reaches Google. A brief built from Search Console describes how people searched Google in the past, and the demand it describes is visible to every competitor through any keyword tool.

Competitor pages. The most explicitly convergent input available. Content built by surveying what rivals cover produces, by construction, the intersection of what rivals cover. It guarantees parity and forecloses differentiation.

General category knowledge. What an experienced writer or a language model already knows about the category. Available to everyone by definition, and the input an AI writing tool defaults to when given nothing else.

The compounding problem is that these are also the four inputs most AI content tools use. A brand writing from public inputs with an AI tool is not competing with four rivals doing the same. It is competing with four rivals running similar tools over the same source material, at higher volume.


3. The four non-public inputs

Four records exist inside most ecommerce businesses that no competitor can read.

Support tickets. Every question a buyer could not answer from your site, in their own words, already answered once by your team. Gartner’s self-service study, published 19 August 2024 with 5,728 respondents, found 43% of customers could not find content relevant to their issue. The tickets are what that failure looks like on your store specifically.

Product reviews. Partly public once published, which weakens them, but the corpus in your review platform includes structured attributes, private feedback and unpublished submissions that a competitor scraping your public reviews does not get.

Returns reasons. A record of buyers who decided with incomplete information. The reason code names the product and the failure mode together. Native data is thin. Shopify’s ReturnReason enum offers a closed list of colour, defective, not as described, size too large, size too small, style, unwanted, wrong item, other and unknown, with returnReasonNote capped at 255 characters and the customer note at 300. Some third-party platforms carry free text: Loop Returns exposes a free-text reason through its warehouse reporting endpoint.

On-site search queries. The closest available proxy for a prompt. Buyers typing in their own words, about your actual catalogue, on a property you own. The highest-value subset is the queries that returned nothing.

None of these is exotic. What is unusual is using them to decide what a page should say.


4. The machine-readability proof

The objection to all of this is practical: interesting in principle, inaccessible in practice. It does not hold. Every source above is readable through documented interfaces today.

RecordAccessNotes
On-site searchShopify Web Pixels API, search_submitted eventCarries event.data.searchResult.query and the returned productVariants. An empty variants array is a zero-result query.
On-site search, reportingShopify Search & Discovery app”Searches with no results” report. Trailing 30 days only.
ReturnsShopify Admin GraphQL, ReturnReason / ReturnLineItemClosed enum plus returnReasonNote (255 chars) and customer note (300 chars).
Returns, free textLoop Returns warehouse reporting endpointFree-text return_reason. It is a warehouse report, not a general returns list.
Support ticketsGorgias, Zendesk, Front, Re:amaze APIsAll expose conversation content and tagging.
ReviewsYotpo, Judge.me, Loox APIsLoox API access is on paid tiers only.

Two honest notes. WooCommerce has no native returns object, so returns data sits in whichever RMA plugin the store runs and access varies. And we were unable to verify Okendo’s API capabilities, because their documentation is client-rendered and could not be read — so no claim is made about it here.

The point of this table is not that the integration is trivial. It is that the barrier is attention, not access.


5. What the evidence says about specific text

Two peer-reviewed results point the same way, and both are worth reading precisely rather than as headlines.

Aggarwal et al., “GEO: Generative Engine Optimization”, KDD 2024. Across 10,000 queries, adding citations, quotations and statistics raised source visibility by roughly 27%. Keyword stuffing lowered it by 8.8%. The widely quoted “up to 40%” figure is a best case for particular categories, not the average, and the study tested general web pages rather than ecommerce product pages.

Filandrianos et al., “Bias Beware”, EMNLP 2025. Across six models and five product sets, social-proof phrasing raised LLM recommendation rates by 9.75% to 22.38%, while scarcity and exclusivity copy cut them by 13% to 46%. Note carefully what was tested: the researchers injected social-proof phrasing, not real customer reviews. It is a proxy, and a counter-intuitive one: the urgency language that works on humans appears to work against you with a model.

Both findings converge on the same conclusion. Specific, evidence-bearing, verifiable text performs. Promotional and vague text does not. First-party records are the cheapest available supply of the former, because a returns log and a support queue are made of nothing but specifics.


6. The provenance audit

Four questions, applied to a brief before writing begins.

  1. Source. Name the record this page draws on. If the answer is “the category” or “competitor pages”, the source is public.
  2. Counterfactual. Could a competitor with only public data produce this page? If yes, it will not differentiate you.
  3. Specificity. Does the page carry a fact, a figure or a quotation that exists only in your records?
  4. Destination. Is this at the right level, product, collection or site-wide? A page at the wrong level fails even with a good source.

A page that fails question two is re-briefed, not deleted. Category pages and buying guides need to exist. The audit does not sort content into keep and bin; it sorts content into what might earn a citation and what will merely be present.

The audit runs in about a minute per brief, and it runs before the expensive part. Every other quality gate in common use, whether similarity scoring, AI detection or editorial review, runs after the page has been written and paid for, and none examine the source at all.


7. What this report does not claim

Several claims circulate in this area that the evidence does not support. We are not making them.

We do not claim AI prefers UGC to brand sites. The largest dataset says the opposite. Any figure showing UGC beating brand-owned domains for citation share should be treated with suspicion.

We do not claim Reddit accounts for around 60% of ChatGPT citations. That figure describes share of prompt responses at a peak, immediately before it fell to roughly 10%.

We do not claim generic content performs worse than publishing nothing. No study exists showing that, in AI retrieval or in search. Ahrefs’ July 2026 analysis of around 331,000 pages found high-AI-signal pages received two to three times fewer impressions and about nine percentage points lower indexation, but that is correlational and rests on a proprietary detector. The defensible claim is that generic content is out-competed and under-indexed, not that it is harmful.

We do not claim Google penalises AI-generated content. The helpful content system was folded into core ranking in March 2024. Google’s spam policy, updated 15 May 2026, targets scaled content abuse and intent to manipulate rankings, not the use of AI as a tool.

We do not claim product reviews measurably change what assistants recommend. No study tests real reviews against commercial assistants. The EMNLP result above is the nearest proxy and it tested injected phrasing.

We do not claim a cost saving from support-ticket deflection. The attribution chain does not close, and the figures in circulation do not survive inspection.

Stating these limits is not throat-clearing. The argument in this report is structural, and a structural argument does not need inflated numbers to stand up. Content written from a source your competitors cannot reach is difficult for them to reproduce. That holds whether or not anyone has yet measured the citation lift.


Conclusion

The question most brands ask about AI discovery is what to write. It is the wrong question, or at least the second one.

Brand-owned pages are the majority of what assistants cite. Ranking has come apart from citation. That combination puts the weight on the content itself, and content written from public inputs is, structurally, content every competitor can produce.

The first question is what the page is made of. Four records inside your business are invisible to competitors, readable through documented APIs, and currently used for operations rather than content. Reading them as content briefs is not a technology problem. It is a matter of deciding that the source of a page matters as much as its quality.

Ask the counterfactual question before you commission anything. It costs a minute, and it is the only gate in common use that runs before the money is spent.

For the companion argument on output rather than input, see The Coherence Report.


References

  1. Profound, “Where do AI citations come from?”, 30 July 2026. 11.84bn citations, 3.02m domains, 16 April to 16 July 2026, 8 models.
  2. Profound with Reddit, AI citation analysis, 10 November 2025. 4bn citations.
  3. Ahrefs, “AI Overview citations from top-10 pages”, 2 March 2026. 863k SERPs, 4m AI Overview URLs.
  4. Ahrefs, AI-content analysis, 27 July 2026. ~331k pages, 100k SERPs. Correlational; proprietary detector.
  5. Aggarwal et al., “GEO: Generative Engine Optimization”, KDD 2024. 10,000 queries.
  6. Filandrianos et al., “Bias Beware”, EMNLP 2025. 6 models, 5 product sets.
  7. Gartner, customer self-service survey, 19 August 2024. n=5,728.
  8. Google Search, spam policies, page updated 15 May 2026.
  9. Shopify, Web Pixels API search_submitted standard event; Search & Discovery analytics; Admin GraphQL ReturnReason.
  10. Loop Returns, warehouse reporting endpoint documentation.

Frequently asked questions

Does this mean we should stop writing category and buying-guide pages?

No. Those pages do a job and buyers use them. The audit sorts content by what might earn a citation, not by what to publish. A page that fails the counterfactual is re-briefed against a first-party source where possible, and published with realistic expectations where not.

We are a small brand with little support volume. Does this still apply?

It applies more, not less. A short list read carefully beats a large keyword export from a different channel. Ten repeated returns against one product is a clear instruction. You are reading these records, not counting them.

Is Search Console useless now?

Not at all. It remains the right tool for Google search performance, indexation and technical diagnostics. The argument is narrower: it is the wrong input for a brief about how buyers ask assistants, because it structurally cannot hold that information.

What is the difference between provenance and coherence?

Coherence tests the output: whether your visible page, your structured data and the rest of your catalogue agree with one another. Provenance tests the input. A page can be perfectly coherent and entirely interchangeable, so both gates are worth running.

Does using AI to write content fail the audit?

No. The audit examines the source, not the tool. A model drafting from your returns log and support queue is working from material no competitor holds. The same model drafting from a competitor's page is not.

How do we start without an integration project?

Export two quarters of returns with reason codes and SKUs, and pull the last thirty days of zero-result searches from your search reporting. Both are available without engineering work, and between them they will name more failing pages than a content audit will.

Next step

Need implementation support after reading?

Translate GEO research into operational first-party pages with Geoffy.

Turn your catalogue into AI-readable product pages and structured data.

Already have an account? Sign in

Structured outputs enabled
First-party pages published
Discovery coverage expanding