The Citation Is Not the Verdict

Most generative engine optimisation is theatre, and the citation data says so. AI engines mostly cite pages that already rank in Google’s top ten, and for a specific business the page they cite most often is that business’s own website. So a citation is not a judgement in your favour. It is the model opening your page to find out what you are. Your copy is the evidence. Nobody is auditing the copy.

Start with the most useful number published all year, because it tells you how much of this industry is invented.

Ahrefs looked at 137,000 sites and found that 97% of llms.txt files never get read (Ahrefs, 24 July 2026, from Louise Linehan’s analysis first published 12 September 2025). Not read late. Not read partially. Never fetched by anything. A file format with conference talks, plugins, agency deliverables and a small consulting economy behind it, sitting in the root directory of thousands of sites, doing nothing at all in 97 cases out of a hundred.

I bring that up not to dunk on llms.txt but because it is a clean example of how this field forms beliefs. Something plausible gets proposed, the plausibility is mistaken for evidence, and eighteen months later there are line items in scopes of work for it.

Meanwhile the actual mechanism turns out to be embarrassingly familiar.

Why do AI engines cite the pages they cite?

Because those pages already rank. That is most of it.

76.1% of URLs cited in AI Overviews also rank in Google’s organic top ten, and citation probability tracks position closely: 33.07% at organic position one, falling to 13.04% at position ten, with pages beyond the top ten roughly four times less likely to be cited (The Digital Bloom, compiled 26 July 2026, citing Ahrefs 2025 and GetPassionFruit 2025). AI Overviews now appear on roughly 48% of tracked queries, up 58% year on year (BrightEdge via The Digital Bloom, February 2026).

So the retrieval layer is not some alien intelligence weighing your brand’s worth. It is largely reading the same ten blue links your competitors have been fighting over since 2009, then summarising them.

And when the question is about one specific business, the answer is even less exotic. Honey Badger SEO collected 1,189 citations across 176 AI-generated local service answers in 22 US markets and found the business’s own website appearing as a cited source three times out of four (Honey Badger SEO, collected August 2026, published 2 September 2026; data page).

Three times out of four. Your own site.

Sit with what that means, because it is the whole post. When a model cites your website, it has not decided you are good. It has gone looking for someone who knows, and the only witness available was you. The citation is the model calling you to the stand.

Is a citation actually a win?

It is an opportunity to be read, which is different.

Here is the failure mode nobody puts in a deck. A model gets asked which agency in Kuala Lumpur handles retail brand work. It retrieves your site because your site ranks. It reads your services page, which says you deliver integrated solutions across the customer journey. It reads your about page, which says you are a passionate team of storytellers. It finds no named clients, no sectors, no scope, no year, no numbers. Then it writes two sentences about you that are technically sourced from your own website and completely useless to the person asking.

You got the citation. You lost the answer.

And you will never see it happen. There is no log of the sentence a model wrote about you, no notification, no complaint. The person asked, got something flat, and moved on to the next name in the list. The only trace is an impression count.

That is not a retrieval problem and no amount of schema fixes it. Structured data tells a machine what kind of thing a page is. It does not supply a fact that is not on the page. If your copy is thin, dated and unspecific, the citation just means the thinness now arrives pre-laundered through an assistant the reader trusts more than your homepage.

There is a second-order version of this that should worry anyone selling visibility dashboards. Microsoft’s 2026 Work Trend Index found that 92% of Malaysian AI users treat AI output as a starting point rather than the final answer (Microsoft, 23 June 2026, 2,000 Malaysian knowledge workers). These are not passive recipients. They read the summary and then go and check. The summary of your page is a trailer for your page. If the trailer is vague, the trailer is what gets remembered.

Why does one citation-count dashboard measure nothing coherent?

Because citation behaviour changes by engine and by question type, so a single count sums things that are not the same thing.

Conductor ran seven months of tracking from September 2025 to March 2026, across seven engines and seven intent categories, 1,056 data points (Conductor, published 7 May 2026). The results do not average well.

Wikipedia was the top-cited domain for ChatGPT Search across Education, Recommendations, Comparison and Purchase intents, and led Education and Recommendations in every one of the seven months. Reddit led Support queries for ChatGPT in five of seven months. Claude, with two months of data in February and March 2026, filled 16 rank-one citation slots and gave zero of them to YouTube, Wikipedia or Reddit. Claude’s top slot went to brand domains and institutional sources.

Read those three sentences as three different worlds, because that is what they are.

One number cannot cover that table. A brand with high citation counts driven by Reddit support threads is in a different situation from a brand with the same count driven by its own product pages. The first is being narrated by its customers. The second is being quoted directly. The dashboard shows the same green arrow for both.

And if you wanted to settle the question with real performance data, the measurement layer will not let you. Google Search Console’s AI performance report gives impressions but not clicks or queries, and it aggregates AI Overviews with AI Mode (Ahrefs, 24 July 2026). You can see that you appeared. You cannot see what was asked or whether anyone moved.

That is the quiet scandal of this whole category. The thing being sold is optimisation against a signal that cannot be closed against outcomes, on a surface whose own analytics deliberately withhold the two fields you would need.

Demand does not care. US monthly search volume for “AI search tracking” is up 184% and “AI rank tracking” up 175% (Ahrefs Keywords Explorer via Ahrefs, 24 July 2026). People are buying instruments for a thing they cannot yet read.

What does this look like from Kuala Lumpur rather than California?

More urgent, and on a shorter clock.

Every citation study above is US-only. The behaviour is not. Deloitte’s Asia Pacific consumer sector work found that nine in ten retail executives expect AI to be used more than traditional search engines by 2026, and half expect today’s multi-step shopping journey to collapse by 2027 (Deloitte, published 14 April 2026, drawing on 185 respondents across Australia, India, Japan and Singapore plus the 2026 Retail Industry Global Outlook dated 8 January 2026). The same work puts agentic AI adoption among Asia Pacific consumer businesses at 29% today, expected to reach 76% within two years.

On the consumer side, 44% of recent Malaysian AI users have used AI embedded in search engines, and 87% have used a chatbot or assistant, out of 67% of Malaysian adults who used an AI tool at all in the past three months (Vodus AI Consumption Study Malaysia 2026, fielded 29 December 2025 to 2 January 2026, 2,556 adults). Usage skews hard by age and income: around 85% of 25 to 34 year olds, and roughly 80% in households at RM8,000 and above versus 48% below RM3,500 (Vodus, 2026). The buyers with money are the ones reading the summary.

Why this matters more here than in the markets the studies sampled: Malaysian brands have thinner English-language third-party footprints. Less trade press, fewer independent reviews, sparser Wikipedia coverage, less of the institutional scaffolding that a model leans on when it wants a second opinion. Which means the Honey Badger pattern bites harder. When there is no encyclopedia entry and no forum thread, the model has exactly one witness, and it is your website.

A US mid-market brand that writes bad copy gets summarised from Wikipedia and Reddit anyway. A Malaysian one gets summarised from its own thin about page. Same engine, worse outcome, and the reason has nothing to do with optimisation.

So what is the actual work?

Being worth quoting when the model opens the page. That is a copy problem, an evidence problem and an editorial problem. It is not a technical one.

I keep coming back to a small test. Take any page on your site, imagine a model has been asked a specific question about your business, and ask whether that page contains a sentence that could be lifted verbatim as the answer. Not a paraphrase of your positioning. A sentence with a fact in it. A sector, a named capability, a scope, a year, a number, a place.

Most brand sites fail that test on every page. Not because they are badly built. Because they were written to feel like something rather than to say something, which worked fine when a human was doing the reading and filling the gaps with a good impression of your design.

Models do not fill gaps with a good impression. They fill them with whatever else they retrieved, or they hedge, or they move on to a competitor whose page had the sentence.

What happens when the summary becomes the whole experience?

The page gets read once, by a machine, on behalf of someone who never sees it.

Deloitte’s Asia Pacific work has half of retail executives expecting today’s multi-step shopping journey to collapse by 2027, and puts agents in a position to influence or handle as much as 25% of global ecommerce sales by 2030 (Deloitte, 14 April 2026). The same data has more than half of Asia Pacific consumer businesses already running live AI implementations across IT, cybersecurity, marketing, sales and customer support, while only around 30% say that at least 40% of their AI initiatives reach production.

That gap between deployment and production is the interesting part, and it rhymes with the llms.txt number. A lot of motion, a thin layer of things that actually work.

But take the direction seriously for a second. If the journey collapses, the page stops being a destination and becomes a source document. Nobody browses it. Something reads it, once, quickly, and reports back. Every design decision made on the assumption that a human would arrive, look around, get a feeling and then find the information becomes dead weight, because the reader is a retrieval pass with no patience and no eye for your typography.

I am not arguing that brand craft stops mattering. I am arguing that the proportion of your site that has to survive being read without a human present just went up, and almost nobody has rewritten for it.

There is a real prize for getting this right, which is worth naming because it is easy to read all of this as pessimism. AI-referred visitors convert at 23 times the rate of traditional organic search visitors (Ahrefs via The Digital Bloom, June 2025), and in one documented case 0.5% of total traffic produced 12.1% of signups over 30 days. The traffic is small and extremely warm. It arrives having already read a summary of you. What decides the conversion is whether that summary was made of anything.

Is any of the technical work worth doing?

Some of it, with the volume turned down.

Crawlability now matters more than it did, because there is far more machine traffic to serve. Cloudflare’s Matthew Prince said in June 2026 that agentic bot, crawler and agent traffic passed 50% of internet traffic for the first time (Ahrefs, 24 July 2026). Blocking the wrong agent, shipping key facts only inside JavaScript or images, and letting your pages sit outside the top ten all have measurable costs.

But notice the order of operations that the data implies. Ranking still drives citation at 76.1%. Citation then hands the model your copy. If the copy has nothing in it, the first two steps were a very expensive way to publish an unhelpful sentence about yourself. Fixing the copy without the ranking is a quiet page nobody retrieves. Fixing the ranking without the copy is what most of the industry is currently doing, at scale, with dashboards.

Stop asking whether the engines cite you. They probably do, because you rank and because you are the only one who has written anything about you. Ask what they found when they got there.

Frequently asked questions

Does llms.txt improve AI search visibility?

There is no evidence that it does. Ahrefs analysed 137,000 sites and found that 97% of llms.txt files are never read by any crawler or agent, so for almost every site the file has no measurable effect.

How do AI engines decide which sources to cite?

Mostly by existing search rank. 76.1% of URLs cited in AI Overviews already rank in Google’s organic top ten, and citation probability falls from 33.07% at position one to 13.04% at position ten.

Is AI citation count a useful metric?

Not on its own. Conductor’s seven-month study across seven engines found citation patterns change by engine and by intent, with Wikipedia leading education and recommendation queries, Reddit leading support queries, and Claude giving its rank-one slots to brand and institutional domains. A single count sums sources that mean different things.

Why does a brand’s own website get cited so often?

Because for questions about one specific business there is often no independent source. Honey Badger SEO recorded the business’s own website as a cited source in three of four local service answers, which means the model is reading the brand’s copy and summarising it directly.