Is AEO Worth It? How to Measure Whether It Pays

AEO is worth paying for when you can read the meter, and most buyers cannot, because the surfaces that quote you do not report back. Before you sign anything, take five readings: share of voice on a fixed prompt set, citations by surface, the words the engines use about you, AI referral sessions, and self-reported first touch. Everything below is how to take them.

One of those five got harder to fake on 31 August 2026, and three of them still have no published benchmark anywhere. I will be specific about which. If you would rather someone else took the readings, our AI visibility audit is £750 and the method is the one written out here.

Is AEO worth paying for?

It depends entirely on what one visitor is worth to you, and there is decent published evidence on both halves of that.

Ahrefs published its own analytics on 16 June 2025. Over 30 days, AI search sent 0.5% of its traffic and 12.1% of its signups. Patrick Stox, who wrote it up, put that at a 23x higher conversion rate than traditional organic search. Ten days later Louise Linehan published a wider Ahrefs study of 81,947 websites: AI referrals were 0.25% of the average site's traffic, roughly ten times what they were a year earlier, while search traffic on those same sites fell about 21%.

Read the two together and the shape is clear. A quarter of a percent of your traffic is doing a wildly disproportionate share of the converting. Stox was straight about the other side of it in the first piece: on Ahrefs data, users from AI search click links 75% less than they do in traditional organic search, so if assistants keep taking share, the absolute number of visits falls even as each visit gets better.

The arithmetic follows from that. If your average deal is worth five figures and your cycle runs months, a thin stream of very warm sessions pays for a retainer quickly. If you sell something cheap on impulse, it does not. Nobody can settle that for you without your own numbers, which is why the measurement matters more than the tactics.

How do you measure whether AEO is working?

Five readings. Take all five on the same day, write the date down, repeat monthly. The table is the part worth stealing.

Two rules separate a measurement from a vibe. Freeze the prompt wording, because changing one word changes the retrieval. And log the model name every time, because an answer from one model version is not comparable to an answer from the next.

What does Search Console tell you now, and what does it leave out?

This changed recently enough that most published advice has not caught up with it. Google announced Search Generative AI performance reports in Search Console on 3 June 2026, and the post now carries a note that as of 31 August 2026 the reports have rolled out to all websites worldwide.

Here is everything the report gives you, taken from Google's own help documentation: impressions, pages, countries, devices and dates. Google defines the metric plainly. "Impressions are how many times links to your site were shown to a user in a generative AI feature on Google Search."

What it withholds matters more:

• No clicks and no click-through rate. Impressions only.
• No position. There is no rank equivalent inside an AI answer, so there is nothing to report.
• No split between AI Overviews and AI Mode. The help page confirms both features feed the same bucket.
• The same 1,000-row limit as the ordinary performance report, so a large site is reading a sample.
• Nothing from Search Labs experiments, which Google excludes explicitly.

Google also notes the data was already inside your overall performance report and has simply been given a view of its own, so your headline totals did not jump on 31 August. If a supplier tells you their work caused a step change in your Search Console numbers that week, they are reading a new chart and calling it a new result.

How do you run the prompt set in practice?

This is the method we run on orrjo.com every month, written out so you can run it without us.

1. Build the list from real questions. Pull them from your sales calls and your inbox. A keyword tool will give you search queries, and search queries are a different animal. They run short and clipped, where the asks that trigger AI answers run long and conversational.
2. Split it three ways. Roughly a third category questions with no brand in them, such as "who does answer engine optimisation for B2B in the UK". A third comparison questions naming competitors. A third questions about you by name. The first group is where you win new buyers. The last group is where you find out what the engines think you are.
3. Run logged out, in a fresh session, from a clean browser profile. Personalisation and memory will hand you a flattering answer if you let them.
4. Record five fields per prompt. Engine, model name, date, named yes or no, and the verbatim sentence. Nothing else, or you will quietly stop doing it by month three.
5. Screenshot anything wrong. A model stating an out-of-date price about you is the most actionable finding you will get, and it vanishes the moment the model updates.
6. Do not change the prompts. Ever. Start a second sheet if you need new ones.
7. Compare on the third reading, not the second. Two points is a line drawn through noise.

I am giving you our method and withholding our readings, deliberately. One agency's share of voice on its own prompt set is unfalsifiable, and publishing it would be marketing wearing a lab coat. What we do publish is work clients will vouch for, and the client testimonials are that with names attached.

What can AEO not do?

It cannot make an AI answer send you a visitor. That is the ceiling. Say it out loud before you buy anything.

Cloudflare put numbers on the gap. In a study published on 1 July 2025, David Belson and Sam Rhea measured crawl-to-refer ratios across their network for the week of 19 to 26 June 2025. Anthropic crawled roughly 70,900 pages for every one visit it referred back. Cloudflare flagged its own caveat in the same post: traffic referred by Claude's native app carries no referrer header, which would overstate that ratio. The direction still holds.

Pew Research Center measured the click side. In a study published on 22 July 2025, Pew analysed the browsing of 900 US adults across 68,879 Google searches in March 2025, of which 12,593 produced an AI summary. Users clicked a traditional result on 8% of visits where a summary appeared, against 15% of visits where none did. They clicked a link inside the summary on 1% of visits. They ended the browsing session on 26% of pages carrying a summary, against 16% of pages without one.

The obvious reading of those two studies is that AI answers are eating clicks, and the obvious reading is only half right. Semrush's AI Overviews study, run across more than 10 million keywords in its US desktop database through 2025, found AI Overviews on 6.49% of queries in January 2025, 24.61% in July and 15.69% in November. Coverage went up, then came back down. Semrush also compared the same keywords before and after an AI Overview started appearing for them, and on those the zero-click rate fell from 33.75% to 31.53%. The trend line has a wobble in it, and anyone selling you a straight line is selling.

So how do you value a citation that sends nobody? Treat it as what it is. A brand impression delivered by a machine your buyer already trusts, with your name in it and your competitor's name absent. Value it at the pipeline layer rather than the session layer, which is exactly what reading five in the table is for. If your enquiries mention an assistant more often than they did last quarter, the citations are working even while your referral chart sits flat.

How long does AEO take to show results?

No primary source publishes a time to first citation. I looked properly. What exists is agency marketing quoting other agency marketing quoting nobody, and anyone who gives you a number in weeks has either made it up or inherited it from someone who did.

What is documented is narrower and more useful. OpenAI's crawler documentation states that after you change your robots.txt, "it can take ~24 hours from a site's robots.txt update for our systems to adjust", and that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links. That is a permission clock rather than a ranking clock, and it is the only published constant I could find that applies to every site equally.

Four things set the real pace, and you can assess all four before spending anything:

• Whether the engine retrieves live or answers from training data. Retrieval-backed answers can change within days. Answers drawn from model weights change when the model does.
• Whether the page exists at all. Most sites lose these answers because nobody ever wrote the page that answers the question, which is a writing problem on a writing timeline.
• Whether anything the engines already trust mentions you. Third-party mentions move faster than anything you publish on your own domain.
• Whether your crawl budget is being spent on the right pages. A technical problem, and normally the cheapest of the four to fix.

Plan on monthly readings and a judgement at reading three. Those are the ranges we plan against, not promises, and any supplier offering a guarantee on an output they do not control is telling you something useful about themselves.

What does an AI visibility audit cost?

Here is what UK providers publish on their own pages, with the date each page carries. Every provider name links to the page the price sits on, so you can run the date check yourself. I have kept this to providers who publish a price at all, because most do not. For the wider picture, including retainers and enterprise programmes, the AEO agency pricing guide is the page that holds it.

Two things fell out of building that table. Half of these pages carry no date at all, which for a discipline that changed materially on 31 August 2026 is a fair thing to hold against a supplier. And the spread between the cheapest monitoring and the cheapest retainer is wide enough to show that AEO pricing covers two different markets wearing one label: monitoring, and production.

If the difference between the disciplines is still fuzzy, SEO, AEO and GEO explained separates them properly, and our AEO service page lists what sits under each heading here.

Where do you start if you are not ready to buy anything?

1. Write 25 prompts this week. Take them from the last ten sales calls. It costs an hour.
2. Run them once, logged out, and record the five fields. That is reading one, and it is the only reading you can never take retrospectively.
3. Open the generative AI report in Search Console. Every property has it now. If your impressions there are near zero, your problem is not measurement.
4. Add the first-touch question to your form today. Every week you delay is a week of attribution you cannot get back.

Do those four and you will know inside a month whether AEO pays at your deal size, which is a far better question than whether it works in general. If you would rather an analyst did it, the £750 AI visibility audit ends with your reading one written down, and the free Website Review on the same page will tell you whether it is worth the £750 before you spend it.

One last test for any supplier you talk to. Ask what reading one looked like for their last client. If they cannot answer, they were never measuring.