Answer Radar

Reading the Mean Position Metric

Mean position answers a question presence and share of voice cannot: when you are named, how early. It is an ordinal, where 1 is the first product an answer names, and it is deliberately the metric with the most refusals attached, because a position is easy to misread.

Quick Answer

Mean position is the average ordinal position at which a vendor first appears across the admissible samples that name it, where 1 is the first product named in the answer. Lower is better. It is computed from the answer text, not from any ranking the model claims, and a vendor nobody named has no position rather than a zero.

  • Per vendor, per segment, over the samples that name the vendor.
  • A vendor named through two product lines contributes one number per answer, its first appearance.
  • Not compared between runs in re-verification, because an ordinal's absence is not zero.

The definition

The average ordinal position at which a vendor first appears across the admissible samples that name it, where 1 is the first product named in the answer. Lower is better. Computed from the answer text, not from the model's own claimed ranking.

Only products are ordered. A review site cited before any product is a citation, not position 1, because publications and features are excluded from the entity set that positions are counted over.

Why the answer text, not the model's ranking

Engines sometimes print a numbered list, sometimes a paragraph, and sometimes a list whose numbering contradicts the order in which brands are actually discussed. Reading the position off the text, first product named is position 1, makes the metric the same shape across engines and across answer formats. A model's claimed ranking is something the answer said; the order it named things is something the answer did.

Why absence is not zero

A vendor no admissible sample named has no position. Reporting 0 would read as 'named before everything else', the opposite of the truth, so the metric is absent for that vendor rather than present as a number. This is the same refusal the ratio layer makes for empty denominators: a value that would be a claim is left out rather than defaulted.

It is also why mean position is not one of the metrics re-verification subtracts. Presence rate and recommendation rate are per-sample ratios on identical denominators, so a difference is meaningful. A mean over the samples that name you has a denominator that changes with presence itself, and subtracting two such means would mix two movements into one number.

How to read it next to the others

Read mean position together with presence rate. High presence and a poor position means you are the also-ran the engine lists after the recommendation. Low presence and a strong position means the engine barely knows you but leads with you when it does, which usually points at one strong cited page. Neither reading is a finding on its own; the gaps and their citations say which pages produced it.

Compare position only within a segment, since each engine formats answers differently, and only across runs on the same instrument, which the disclosure block identifies.

  • Position 1 to 2 with high recommendation rate: you are the answer.
  • Position 4 or worse with high presence: named, not chosen. Look at who is cited first.
  • No position: no admissible sample named you. That is a presence problem, not a position one.

FAQ

Is a lower mean position always better?

Yes, 1 is first. But a low position over two samples is a weaker claim than a slightly higher one over twenty. The disclosure block shows how many samples named you.

Why does my position change between runs when nothing changed?

Answers are not stable. The same question asked twice can order brands differently. A series over several runs, on one instrument, is the honest way to read position; a single run is one sample of it.

Does a mention in a comparison table count as a position?

Position is the order of first appearance in the answer text as recorded. A brand named first in a table the engine returned as text is first. How each engine's formatting is captured is part of the instrument the disclosure names.

Related help pages