Reading an answer

Your assistant writes the answer you read, from numbers Hiasynth computed. Those numbers have a small vocabulary: share, lift, baseline, index, confidence. Knowing it lets you judge an answer and ask a sharper follow-up.

The shape of every answer

Each tool answers in the same order, most important first.

• A headline. One sentence with the takeaway, written on our side from the numbers.
• The size. A count, the population it was counted in, and the share.
• What stands out. Composition, places, channels or values, depending on the tool.
• The baseline everything was compared against.
• A confidence tier and the data vintage.

Every share, lift, index and rank is calculated by the server. Your assistant quotes them and does not do the arithmetic, which is where language models are least reliable.

Share and lift

A share is a plain percentage: 84.7% of this audience work full time.

A lift is that share divided by the same share in the baseline: 84.7% against 49% is 1.73×. Above 1 means more common than expected, below 1 means less. A lift of 0.14 on "divorced" says as much about an audience as a lift of 1.95 on "informed digital media diet".

Answers give both, and good answers quote both. The share is what a reader can check. The lift is what makes it a finding.

full-time worker   84.7%  vs 49.0%   1.73×
married            79.2%  vs 49.0%   1.62×
divorced            1.3%  vs  9.9%   0.14×

The composition of affluent 30 to 39 year olds in the Netherlands, against everyone in the Netherlands.

The baseline

A lift means nothing without knowing what it is a lift over, so every answer names its baseline.

• In analyze_market and reach_audience it is everyone else in the place you asked about: "within the Netherlands".
• In describe_area it is the rest of the country: "the rest of SE".
• In a lifestyle block, in describe_area and in a persona, it is the national average for households: "the SE average". A figure of 1.4 on eating out means 1.4 times the share of the budget a typical household in that country gives it.
• In compare_places it is the places you named, pooled together, and not Europe.
• A whole country described on its own has no baseline, and the answer says so instead of inventing one.

"1.5× the national rate" and "1.5× comparable cities" are different claims. If you want a different comparison, ask for it.

Core markets and hotspots

analyze_market returns two lists of places, and they answer different questions.

Core markets are where the most of your audience live. They are nearly always the big cities, because big cities have the most of everyone.

Hotspots are where your audience is most concentrated, by lift. This is where a campaign wastes least and where a shop has the densest catchment.

Amsterdam is on both lists for affluent thirty-somethings. Diemen, with under two thousand of them, is only on the second. Whether Diemen matters depends on whether you are buying national media or choosing a street.

Each place shows its count, the population it is out of, the local share and the lift. The population is the denominator: treat a large lift on a small denominator with care, even though the server has already been conservative with it (see below).

Reach, density and fit

rank_places scores every candidate twice. Reach is how many of your audience live there. Density is how concentrated they are against the surrounding rate. Both are scaled against the best candidate, and fit combines them with the weight you chose: 0 is all reach, 1 is all density, 0.5 is the default.

The answer prints the formula and the weights it used. It also tells you when the choice mattered: "Stockholm has the most people, Mörbylånga the highest concentration." If that sentence appears, decide which you care about and ask again with the weight set.

Index against the pool

compare_places gives each place an index against all the compared places pooled. 1.10 means the audience is ten percent more common there than across the set. Each place also gets two ranks, by size and by density, and the headline says when they disagree.

Values in words

Attitudes and values come from the European Social Survey and are stored as z-scores: 0 is the European average, positive is above, negative is below. The tools translate them:

environmentalism         +1.15   notably above average
open-mindedness          +0.50   notably above average
wealth and status        −0.36   notably below average

Read the negatives as carefully as the positives. "Notably below average on wealth and status" tells you what tone to avoid. When a place or an audience has no distinctive values, the list comes back empty and the headline says the profile is close to average. That is a finding too.

Money

Amounts ending in _tlc are thousands of the local currency. A median net income of 297.6 in Stockholm is 297,600 kronor. Within a country they compare freely. Across countries they don't, and the answers use quartiles and deciles instead.

Confidence

Every answer carries a tier.

Two more signals are worth noticing. analyze_market says how each part was computed: the count is always a full count, while the composition of a very large audience may be estimated from a random sample, and says so. And a place answer may carry a coverage note when part of a region is still being built.

Small numbers are pulled back

A square kilometre with forty matching people out of a hundred would report a spectacular lift and mean nothing. Hiasynth shrinks every local rate toward the rate of its surroundings, in proportion to how little evidence there is, before it computes a lift. A place needs roughly a thousand people before its own rate counts fully.

So the lifts you see are already conservative. A small place that still shows up as a hotspot has earned it. Don't let your assistant re-inflate it, and don't read a modest lift on a small base as a weak result.

When nothing stands out

The notable finding is only reported when it is material, a concentration of at least 1.25× on a solid base. When nothing clears that bar, the answer says so plainly. An honest "this audience is spread evenly" saves you from a geographic strategy you didn't need, and it is what makes the findings that do appear believable.

When the answer is no

Suppressed. Anything under 100 people is not shown. The answer comes with a recovery line that says which lever to pull: "248,016 people live in Uppsala and 12 match these filters. The area is fine; the segment is what's thin." Or the opposite, when the place itself is tiny.

No rows matched. Zero is different from thin. It almost always means a value that doesn't exist, such as Q5 on a quartile, and the answer says to check the values instead of widening anything.

Unknown column. The answer suggests the closest real attributes and says which table they are in.

Coarsened. A sensitive attribute at fine geography, or a place too small to describe, is answered for the surrounding region, with a note saying that it was.

In every case the message is written for your assistant to act on, so the usual result is that it corrects itself and asks again.

Next

Accuracy covers how the population is calibrated and where it runs thin. The nine tools lists what each tool returns.