How we work
The right questions first. Then a measurement you can trust.
Aions works in four steps: define the questions, measure, audit and improve, measure again. This page explains how that works and where the limits lie. The technical details are further down.
No account, no subscription, no cost.
The process
Four steps, always in this order.
- 01 Definewhich questions decide in your category and on which AI systems we measure.
- 02 Measurewho is mentioned and recommended for these questions, every day, with disclosed precision.
- 03 Audit and improvewhat AI agents see on your pages and what needs to change, with acceptance criteria.
- 04 Measure againthe same questions, the same systems, after implementation.
01 · Define
A measurement can be precise and still say nothing.
If you only measure the questions you would like to be found for, you get precise figures for the wrong question. That is why we measure decision questions: the questions buyers ask shortly before a purchase, with conditions such as price, delivery time or returns.
- Core questions stay the same for months. That is the only way to compare.
- Exploratory questions test new ideas, competitors and variants. They are reported separately.
- You can add your own questions at any time.
Representative data on how people ask AI systems hardly exists today. We therefore disclose the questions, give our reasons for them and review them continuously.
| Question | Condition | Mention Rate | Chosen instead of you |
|---|---|---|---|
| Which office chair helps with back pain? | none | 41% · 49 of 12032 to 50% | You lead |
| Office chair under CHF 600, delivered by Friday? | price, delivery time | 6% · 7 of 1203 to 12% | Provider A |
| Where can I try a chair for 30 days and return it free? | returns | 0% · 0 of 1200 to 3% | Provider C |
| Ergonomic office chair made in Switzerland? | origin exploratory | 6% · 1 of 161 to 28% | Provider B |
Last 30 days · core questions daily (120 answers per question), exploratory questions weekly (16 answers)
Three different return periods on your pages
Finding CONFLICTING_RETURN_POLICY · action ready
02 · Measure
AI answers fluctuate. So how often you measure matters.
The same question does not produce the same answer every time. A single answer is a sample, not a measurement.
A worked example: at an assumed rate of 10%, a month of daily measurement narrows the value down to 3 to 26%. Weekly measurement only narrows it down to 1 to 58%.
Our recommendation: core questions daily, exploratory questions weekly, four weeks daily after an action. We only highlight changes once the intervals no longer overlap.
95% Wilson interval, one question on one system.
The last row combines 25 questions: 750 answers instead of 30, hence the narrower interval. It applies to the average across all 25 questions; each single question stays at 3 to 26%.
The narrower the bar, the more precise the figure. Repeated answers are correlated, so the actual interval is wider.
03 · Audit
We fetch your pages with the user agent of each AI service.
AI services fetch pages with their own user agents: for the search index, for training or live when someone asks. These fetches can receive entirely different content.
That is why we fetch your pages with each user agent individually. A fetch with a normal browser does not show what a crawler gets. In one of our projects, the index crawlers received the content, while the assistants' live fetches received only an empty page.
Every finding comes with an action, an acceptance criterion and a responsible party.
04 · Measure again
The same measurement before and after. Whoever implements.
After an action, we measure with the same questions, the same systems and the same rules. We recommend four weeks of daily measurement. A re-test in the Partner programme runs for the same duration as the first measurement. The daily trend shows from when answers change.
Aions does not implement itself. Implementation is done by your team, your agency or an implementation partner. The measurement is always the same; denominator and interval stay visible. We also report when nothing changes.
- Mentioned
- in 17 of 25core questions
- Cited
- 11%330 of 3,000 · 10 to 12%
- Chosen first
- 7%63 of 900 with a clear choice · 6 to 9%
- Average position when mentioned
- 3.6in 540 answers with a mention
- Where links lead
- 58% lead to a page that takes an order238 of 410 links · 53 to 63%
14% (12 to 16%) after 21% (19 to 23%)
21 Sep to 2 Oct · 3 to 20 Oct · action 3 Oct · Intervals do not overlap- 1Provider A14% 1,917
- 2Provider B12% 1,598
- 3Provider C10% 1,364
- 4Provider D9% 1,201
- 5Provider E8% 1,066
- 6Provider F7% 958
- 7Provider G6% 823
- 8Provider H5% 709
- … 2 brands in between
- 11Your shop4% 540
18% your shop · 17 to 19% · 540 of 3,00051% average of 4 competitors
- Your shop
- Provider A
- Provider B
- Provider C
- Provider D
- Action 3 Oct
- Your shop
- Average of 4 competitors
- Action 3 Oct
scale 0 to 70% · band: 95% interval, your shop
11% your shop · 10 to 12% · 330 of 3,00010% average of 4 competitors
- Your shop
- Provider A
- Provider B
- Provider C
- Provider D
- Action 3 Oct
- Your shop
- Average of 4 competitors
- Action 3 Oct
scale 0 to 20% · band: 95% interval, your shop
4% your shop · 540 of 13,50011% average of 4 competitors
- Your shop
- Provider A
- Provider B
- Provider C
- Provider D
- Action 3 Oct
- Your shop
- Average of 4 competitors
- Action 3 Oct
scale 0 to 16%
For specialists
Metrics, statistics and audit rules in detail.
Metrics
- Mention Rate: share of answers in which your brand appears in the visible text.
- Recommended: share of all answers that recommend your brand (not just mention it).
- Win Rate: share of answers with a clear choice in which your brand is chosen, across the core questions in a category. Exploratory questions are not included. Shown in the dashboard as „Chosen first“.
- Where links lead: share of links in the answers that lead to a page that takes an order. Links are counted, not visits.
- Citation Rate: share of answers that use your domain as a source, whether or not you are mentioned.
- Share of Voice: your share of all mentions in the field.
- Average position when mentioned: your place within an answer when you are mentioned.
- Rank in Share of Voice: your place among all brands in the category. The dashboard shows the two positions separately.
Statistics
- Every rate with denominator and 95% Wilson interval.
- A change is only highlighted once the intervals no longer overlap. This is stricter than a simple test of the difference.
- A re-test measures with the same questions, the same systems and the same duration as the first measurement.
- Answers on the same day are more closely correlated. One run per day is therefore more informative than three runs every three days.
- Core questions are the stable denominator. New core questions are only added at a documented cut-off, with the break in the series shown. Exploratory questions are never mixed into trend values.
Measurement
- Queries per system via API or through the chat interface. Model and access path are documented for each run.
- Results kept separate for each system, combined only if you choose to.
- Open competitive field: we do not give the model any competitors.
- Fan-out queries are logged wherever a system discloses them.
Readiness audit
- Register of official crawler user agents with variants, matrix page × user agent by role: index, training, live fetch.
- Markup and infrastructure audit on the best available fetch; „blocked by delivery“ is a finding in its own right.
- Up to 100 pages: the main language with a sample per page type and template, other languages with the pages on purchase conditions. Googlebot and Bingbot on a sample only (home pages, one page per template and language, all pages with conditions). We fetch robots.txt, llms.txt and sitemaps once with a browser user agent and evaluate the rules for each user agent; the pages themselves with each user agent separately. Pages not in the sitemap but linked from the home page are added; the report lists them.
- Our evaluation is based on what the AI crawlers receive. Only if no AI user agent received the content of a page do we evaluate the response to Googlebot or Bingbot, and we mark this in the evidence.
- Every finding is re-checked live before the report is sent and reviewed by us.
Limits of the measurement
- Without a comparison group, we do not say why something has changed. We show whether it has, and by how much.
- We do not know your customers' actual questions, nor any search volumes in AI systems. We derive the questions and disclose them.
- We do not promise a specific position or a specific Win Rate.
- We query chat interfaces with a Swiss location and without logging in. Logged-in users may see different answers.
- We fetch your pages under the name of each crawler. Some websites also recognise real crawlers by where they come from; our requests do not replicate this. That is why we also fetch selected pages through an AI provider's infrastructure (currently Anthropic). For other providers, only your server logs show this. We only report the fetch through Anthropic's infrastructure as a finding if several pages fail to arrive in two runs; errors of the tool itself do not count.
- Blocks that only affect our test request, search engines included, are not reported as findings. If the server blocks AI crawlers while search engines receive the page, that is a finding. If only training crawlers are refused, we mention it as a note.
- We do not bypass blocks: we do not solve bot-protection challenges and do not switch IP addresses to obtain content. What an AI assistant cannot receive from you is reported as a finding.
- We audit the „purchasable“ stage as preparation. We do not claim any link with recommendations as long as there is no data.
See what this looks like for you.
The free check shows up to three of the most serious findings on your website.
No account, no subscription, no cost.