What have the authority sites measured recently?

On 2026-10-02, Ahrefs published “AI Search ROI Is Messy. Here’s How to Measure It Anyway.” The title says it plainly: the return on AI search really is messy, but they still recommend measuring it. Reading that, I felt relieved, because it matches my own experience — messy does not mean unmeasurable; messy means you cannot measure it the old way.

Their approach is to break “the results AI search brings” apart instead of forcing out one overall ROI figure. That matters a great deal to export-business owners. When you ran Google Ads in the past, you could at least line up a click with a conversion; now an AI answer reads your website’s content and mentions it once, and the user may never click through at all — or may look at three other vendors before coming to you. In your analytics you see only a small wobble in organic traffic, with no way to tell whether AI contributed a share of it.

Semrush took another route in “How to build an AI visibility dashboard” on 2026-10-12: first turn AI visibility into a dashboard you can look at every day. The two firms head in different directions, but they share one point — accept that metrics must be layered before you talk about integration. I make the same split in how to define AI visibility monitoring; I just use a spreadsheet instead of a tool.

Why can’t a citation be counted directly as ROI?

Suppose ChatGPT cites your article once. You are pleased, and you put “12 AI citations this month” into your report. Then the question comes: of those 12, how many point to your own website, how many point to an article you published somewhere else, and how many merely mention a product name? If all three are mixed together, that 12 is just a number that sounds good.

In “How ChatGPT Does Query Fan-Out (15 Million Queries Studied)” on 2026-10-12, Ahrefs lays out query decomposition: a user asks one question, and the model may split it into several sub-queries to search behind the scenes. That means a single answer may have read a dozen pages at once. Of the pages that get read, only a few are written into the answer, and of those, only a few carry a link. Treating “was read” as “was cited” is the most common source of inflated numbers.

There is another layer below that: being cited does not equal being visited. Users read the answer and leave, which is entirely normal with AI answers. That is why, internally, I insist on recording mentions, citation links, site visits, and inquiries separately. You can check Search Console’s AI-related query report and a breakdown of GA4 organic traffic to join the platform-side records with the site-side records. Where they cannot be joined, leave a blank rather than papering over it with one composite score.

From citation to inquiry, which entries have to be recorded in between?

I stretch this into four layers. Each layer has its own evidence, its own owner, and its own most common misunderstanding.

The place most likely to break is between the second and third layers. You can still check the content and entry layers by hand; the visit layer starts to depend on analytics, and the business layer depends on people. A common situation at export sites: the content layer changes, the business layer records no source at all, and in the end nobody can answer “how many inquiries did AI search actually bring?” It helps to settle what a monthly monitoring report should include before discussing a bigger budget.

The homeliest method for the business layer is also the most effective: add a “How did you find us?” dropdown to your form, and make the first question on WhatsApp and email “Where did you hear about us?” Do not underestimate that one line; it is more reliable than any attribution model, because it asks about the customer’s own memory. The cost is a single exchange.

What did our own panel actually measure?

Back to my own records. The T0 baseline round was 2026-09-17 and the T1 monitoring round was 2026-10-12; prompts were fixed by topic, language, and phrasing, with 72 answers recorded per round. Strict brand mentions went from 3/72 at T0 to 6/72 at T1; on vendor-type questions, appearances on a shortlist or in a category table went from 0 to 2. Within the same rounds, the total number of cited domains moved from 322 to 245, and only 90 appeared in both.

Those three numbers only mean something together. The doubling of mentions is real, but it happens at the answer level; the low domain overlap shows that the source pool itself is rotating hard. If I reported only the first sentence, it would be easy for the boss to think “AI is already recommending us consistently.” The fact is that the same kind of question pulls in very different pages from round to round, and we have merely edged up a little on a very unstable line.

Here is a colder example: of the seven canary pages, only by D+14 did all 7/7 get indexed, and only 2/7 earned impressions. The site’s overall indexation rate is 100%, yet in 28 days it had just 12 impressions and 0 clicks. This is not proof that “AI search does not work”; it only reminds me that indexation, impressions, clicks, and citations are four different curves, each moving at its own speed. I keep the related method records in field research and method notes, so judgments and raw data can be reviewed separately.

How do you build a bookkeeping sheet that costs nothing?

No need to buy a tool. One spreadsheet with four columns is enough to start running today. Column one is the date, column two is the exact prompt text, column three is the evidence about you that appears in the answer (mention, link, and name check, written on three separate lines), and column four is what changed on your pages during that period. Run it every two weeks with the same fixed prompts.

The value of this sheet is not that it looks good; it is that it can answer “which layer broke?” for you. If column three barely moves for two quarters straight, the content layer is not working — leave visits and business aside for now. If column three is rising while the visit layer stays flat, the focus shifts to the entry layer: is the link missing from the answer, or is it there but the landing page fails to catch the visitor? If the visit layer moves but inquiries do not, the problem is most likely in the page and the sales pitch, not in AI.

Whatever you do, do not skip the last column. Without “what we changed during this period,” you cannot explain the swings in column three; you will simply credit yourself for every rise and blame the algorithm for every fall. Check it alongside which metrics in a report are vanity metrics so you do not treat the numbers in your panel as achievements outright.

A low-cost action checklist

Just three steps, doable this week, at zero total cost.

1. Set 20 fixed prompts: based on your main products, write them as sentences a buyer would actually ask (how to choose, where the price differences are, how long lead times run), keep one set in Chinese and one in English, and use the same batch every round from then on.
2. Keep the four-column ledger: date, prompt, the evidence about you in the answer, and page changes during the same period. Use one shared spreadsheet that anyone can fill in.
3. Add one source-confirmation question to inquiries: a dropdown on the form, plus a manual question in chat and email. Without this at the business layer, the first three layers can look as good as they like and the records are still wasted.

These three steps do not settle “what AI search is worth.” They settle “whether you are in a position to calculate it at all.” The Ahrefs article of 2026-10-02 offers the measurement method; I offer the version that fits into an export company’s daily routine. If you want to start doing, continue with where to start with AI search optimization.

The questions owners ask most often

Can AI search ROI actually be calculated?

Yes, but not with a single number. Citations, name mentions, site visits, and inquiries have to be recorded as four separate layers, each with its own evidence, before you can judge which layer has broken.

If our company name appears in an AI answer, does that count as a result?

It counts only as one observation at the content layer. A name appearing does not mean anyone clicked through, let alone sent an inquiry; you still need visit records and sales confirmation.

With no budget for monitoring tools, what can we still record?

Record three things first: the fixed prompts, screenshots of the answer text, and when your pages changed and what changed. All three fit in a single spreadsheet and depend on no paid tool.

How long does it take to tell whether AI search is worth investing in?

I suggest reviewing by quarter rather than by week, because indexation, impressions, and citations all lag. Run the fixed prompts for a full quarter first, then decide whether to add budget.

If citations and inquiries rise at the same time, can we say AI search caused it?

You cannot conclude that directly. You still have to rule out same-period ads, trade shows, repeat purchases by existing customers, and other factors; at minimum you need one record that links a source to an inquiry.