Your monthly reporting template has a column headed "Position." It has worked for a decade. Now you are being asked to add AI assistants to the same sheet, and there is nothing honest to put in that column.
Do not fill it in. The column is the problem, and the fix is to retire it for this surface and replace it with metrics that describe what an answer actually is. This post does both in order: why the question feels right, why it does not map, and exactly what goes in the template instead.
Why "What's Our Position" Feels Like The Right Question
Because for search results it was a good question, and for good reasons.
A results page is an ordered list produced by a ranking function. The ordering is stable enough to sample, the positions are discrete and countable, position correlates with attention in ways that have been studied for years, and the whole industry built shared vocabulary and tooling on top of that fact. Asking for position was asking for a number that was well-defined, comparable over time, and connected to an outcome.
None of that is wrong, and traditional rank tracking has not become worthless. Classic search results still exist, still drive clicks, and still deserve the reporting they get. The mistake is not caring about position. The mistake is assuming the concept transfers to a surface built on a different mechanism.
What An Answer Actually Is, Structurally
An assistant's answer is generated text. It is composed token by token, conditioned on the prompt, on whatever the system retrieved if it retrieved anything, and on the model's parameters. There is no ordered candidate list underneath it that got truncated for display. That structural fact has three consequences that matter for measurement.
There is no fixed slate. A results page has ten organic positions whether or not anything relevant exists. An answer may name one brand, five brands, or none, and may name a category of solution without naming a company at all. The denominator is not fixed, so a "position out of N" has no stable N.
The output varies between runs. The same prompt run twice can produce different brands in a different order. A position measured once is a draw from a distribution, not a coordinate. The mechanics behind that variability are covered in why answers vary between people.
Order emerges from generation, not from scoring. When a model lists three options, it is producing fluent text, and the sequence is shaped by phrasing, sentence construction, and what came before in the response as much as by anything resembling a comparative judgement about the brands. Put those together and the request "tell me our position in ChatGPT" is asking for a coordinate in a space that has no coordinates.
Order Of Mention: What It Does And Doesn't Indicate
Being named first is not nothing. It is just much weaker evidence than it looks, and it should never be reported as a rank.
What order of mention plausibly indicates, in aggregate and over many runs: that a brand is strongly associated with the phrasing of the prompt. If one name appears first in a large share of runs across varied phrasings, that is a real pattern worth noting.
What it does not indicate:
- A quality judgement by the model. Being named second does not mean the system rated you second.
- A stable rank. Run the prompt again and the order may change with nothing else different.
- A specific outcome. There is no established relationship between order of mention in a generated answer and any downstream commercial result, and this site will not assert one. Anyone quoting a conversion difference by mention position is quoting a number without a published method behind it.
- Comparability across prompts. First place in an answer that names two brands and first place in an answer that names eight are not the same observation.
The practical rule: treat order as a secondary, aggregate signal with wide error bars, reported alongside a sample size, or do not report it at all. Never as a single integer per month. Never with an arrow next to it.
There is a related trap next door. A brand being mentioned is not the same as a brand being recommended, and an answer that names you as an option to avoid still increments a naive mention count. That distinction gets its own treatment in when a mention isn't a recommendation.
The Metrics That Replace Position
Four metrics do the job position used to do, each answering a narrower question honestly. Formulas, worked examples, and the specific failure mode of each are in the companion guide on which metric answers which question; what follows is the mapping.
| Old question | Old metric | What to use instead | What it actually answers |
|---|---|---|---|
| Are we visible for this query? | Ranking position | Mention rate, with sample size | The proportion of runs of a prompt in which the brand is named at all |
| How much of the results page do we own? | Share of the top 10 | Share of voice across a prompt set | How much of the named-brand space in a category's answers is yours, relative to competitors |
| Which of our pages is ranking? | Ranking URL | Citation share | Which sources the system surfaced, and how often yours were among them |
| Are we shown favourably? | Not really tracked | Sentiment and framing of the mention | Whether being named helped or hurt |
| Did we move? | Position change | Change against a baseline, with variance stated | Whether an observed difference exceeds what run-to-run noise would produce anyway |
Two structural changes come with this mapping. First, every one of these is a rate or a proportion, so every one needs an explicit denominator and an explicit sample size or it means nothing. Second, all of them are prompt-set-level metrics rather than single-query metrics. You measure a set of representative buyer questions, not one keyword, and your reporting unit is the set.
How To Rewrite Your Reporting Template
A workable replacement, in the order the columns should read:
- Surface — which assistant, which mode, which access route.
- Prompt set version — the frozen, versioned set the numbers came from.
- Runs per prompt — the sample size, per cell, visible in the table rather than in a footnote nobody reads.
- Run dates and model version — where the API exposes it.
- Mention rate — with a stated range, not a bare point estimate.
- Share of voice — with the denominator named in the column header.
- Citation share — only for surfaces where citations are near-universal; leave it empty rather than filling it with a number built on a handful of cited runs.
- Sentiment or framing — even a coarse three-level classification beats omitting it.
- Change vs. baseline — flagged as material only where it exceeds observed variability.
And the deletions: the Position column, any per-prompt integer rank, and any single blended "AI visibility score" that rolls surfaces together without published weights.
The most likely way this goes wrong is not refusing to change. It is changing the labels and keeping the habit. A "share of voice" reported as 41% with no denominator, no sample size, and a green arrow beside it is the position column wearing a new hat, and it will be defended in a meeting with the same false confidence.
Guard against that with three rules. Never report a rate without its sample size in the same visual unit. Never report a change without stating the variability you would expect from re-running the same measurement unchanged. Never let a metric acquire more decimal places than your sample can support — a figure computed from a small number of runs does not become precise by being written to one decimal place.
If you want the underlying discipline rather than just the template, how to measure AI visibility sets out the sampling and provenance requirements these columns depend on, and the glossary has the quotable one-line definition for each metric for when someone asks mid-presentation.
In Plain English
Assistants do not rank. They generate text, and the order brands appear in comes out of that generation rather than out of a scored list, so there is no position to report. Being named first is weak aggregate evidence, not a rank, and it should never be reported as an integer with an arrow next to it.
This matters to anyone maintaining an SEO reporting template that now has to cover AI answers, and to anyone whose leadership is asking for a single number. Use this when you are rebuilding the template. Do not use it as an argument to drop classic rank tracking, which still measures a surface that still exists.
Next step: delete the Position column for assistant surfaces, add mention rate with a visible sample size, and make sure every rate in the sheet has a denominator written in the header.
Get the formulas, worked examples, and failure modes for each replacement metric in the metrics guide before you commit to a primary number.
Primary-source verification
Checked July 28, 2026. The operational and factual guidance above was reviewed against the primary references below. Provider features, prices, policies, interfaces, jurisdictions, and account-specific results can change; recheck the controlling first-party document and your own account or measured data before acting.
- Google Search: AI features and your website - How Google describes eligibility, preview controls, and appearance in its AI search features.
- Google: common crawlers and special-case user agents - Current Googlebot and Google-Extended identities and controls.
- OpenAI: overview of OpenAI crawlers - Current GPTBot, OAI-SearchBot, and ChatGPT-User descriptions and publisher controls.
- Anthropic: web-crawler controls - Current ClaudeBot, Claude-SearchBot, and Claude-User descriptions and controls.
- Perplexity: crawler and user-agent documentation - Current PerplexityBot and Perplexity-User descriptions and controls.