Key takeaways
- →An anchor is any value the group encounters before it judges: the first estimate spoken, the number already on the slide, or the option printed at the top of the list.
- →Anchors work even when the group knows the number is arbitrary and even when the facilitator warns them about it, which is why warnings are not a fix.
- →Group votes are worth more than individual ones only when the judgements are independent; anchoring is the fastest way to destroy that independence.
- →The reliable counter-measure is sequence, not wording: silent individual estimates first, simultaneous reveal second, discussion third, one re-vote last.
- →Show the spread of a first-round vote rather than the average, because an average broadcast back to the room becomes the anchor for round two.
- →A deliberate anchor is legitimate when you are calibrating a group to a real constraint, as long as you say out loud that it is a constraint and not an estimate.
The anchoring effect voting problem is easy to state and very hard to see while it is happening: whichever number, name or option a group meets first drags every judgement that follows towards it. A room that would have produced honest estimates ranging from two weeks to four months will cluster within a fortnight of whatever the first person says out loud.
This is not a failure of nerve or a sign of a weak team. Anchoring works on people who know the anchor is arbitrary, on experts judging their own field, and on facilitators who have just finished explaining the bias to the room. Being told about it changes almost nothing. Changing the order in which people speak changes almost everything.
What follows is a working guide rather than a literature review: where anchors get into a session, how to spot an anchored vote from the shape of the results, an eleven-minute sequence that produces an unanchored first round, and what to do when you are halfway through a vote and realise the damage is already done.
What an anchor actually does to a vote
Anchoring is an adjustment failure. People take the first plausible value they are given and adjust away from it, and they consistently stop adjusting too early. The result is a set of answers that look like independent opinions but are really one opinion plus a scatter of small corrections around it.
That matters more in groups than alone, because the entire case for asking twelve people instead of one rests on their errors being independent. When estimates are genuinely independent, the aggregate is usually better than most of the individuals in it. When they are all adjustments from the same starting point, you have not collected twelve judgements. You have collected one judgement and eleven echoes, and averaging echoes does not cancel anything out.
The practical consequence is that a beautifully facilitated discussion can still produce a worse decision than a scrappy silent poll. Discussion first, vote second is the default in most organisations, and it is the exact ordering that guarantees the loudest early number sets the range.
Where anchors get into a session
Anchors are rarely introduced deliberately. They arrive in the framing, the slide deck, the option order, and the person who happens to be least afraid of speaking first. If you want to find them, run through the four minutes before your last vote and ask what number or name was visible.
The sequence matters more than the wording
Most attempts to fix bias in group voting go after the question wording. That is worth doing, but it is second-order. You can write a perfectly neutral question and still get an anchored result if you ask it out loud, in a room, after someone has already floated a figure.
The two structural techniques that reliably help are silent individual generation before any group exchange, and simultaneous rather than sequential reveal. Nominal group technique and the Delphi method are both built on that principle: get the private judgements out of people's heads and recorded before anybody sees anybody else's.
In a live session, this is the difference between asking for a show of hands and running a poll where nobody sees the tally until submissions close. The activity looks near-identical from the outside. The information you get back is not comparable.
An eleven-minute debiased voting sequence
Frame without a number (60 seconds)
State what is being decided and what a good answer would look like, with no example figure, no range, and no reference to what was decided last time. If someone asks for a ballpark to get started, say that you will share the comparison after the first round.
Silent individual answer (3 minutes)
Everyone writes or submits their own estimate or choice privately. Three minutes is enough for a sizing question and generous for a multiple-choice vote; the point is that nobody is waiting on anybody else, so nobody is copying.
Close and reveal the spread, not the average (2 minutes)
Show the full distribution: highest, lowest, and where the mass sits. Do not announce a mean. The moment you say "the average was fourteen", you have handed round two a fresh anchor stronger than the one you just avoided.
Interrogate the extremes first (3 minutes)
Ask the highest and lowest responders what they know that the middle does not. This is where the real information lives, and starting there stops the discussion collapsing towards the mode in the first thirty seconds.
One re-vote, then stop (2 minutes)
Run the same question once more, silently. If the spread narrows, you have learned something. If it collapses to a single value, you have watched the room converge socially rather than genuinely, and that is worth naming before you act on the number.
Wording fixes that cost you nothing
Top Tips
- Strip comparison figures out of the question itself. "How long will this take?" collects better data than "Will this take more or less than six weeks?", which has already told the room to think in weeks.
- Randomise option order between participants where your tooling allows it, or at minimum reorder the list between the first and second rounds so the same item is not permanently sitting at the top.
- Ask for a range rather than a point estimate on sizing questions. People defend a point; they will report the width of their uncertainty far more honestly.
- Ask the least senior person in the room first when you do run an open round, and say why you are doing it. The explanation is not a courtesy, it is what stops the reordering reading as a test.
- Avoid round numbers in any example you must give. A room anchored on 100 rounds to 100; a room anchored on 87 at least has to think.
- Where a historical figure genuinely matters, show it after the first round as a calibration check rather than before it as context.
Anchoring, bandwagon and groupthink are three different failures
These get used interchangeably and they need different fixes. Anchoring is about a value that arrived first. The bandwagon effect is about a majority that became visible. Groupthink is about a group that has stopped wanting to disagree with itself. You can eliminate one and still be badly exposed to the other two.
| Failure | What triggers it | The fix that actually works |
|---|---|---|
| Anchoring | A number, name or option encountered before people judge | Silent independent answers before any exchange; reveal the spread, never the average |
| Bandwagon | A running tally or early majority visible while voting is open | Hide live results until submissions close; reveal everything at once |
| Groupthink | A cohesive group under time pressure with a preferred answer | Assign a dissent role, run a pre-mortem, take anonymous input before the meeting |
| Social loafing | Individual contribution invisible inside a large group | Smaller units, named or counted contributions, shorter rounds |
Reading an anchored result from the shape of the data
You do not need to have witnessed the anchor to detect it. Anchored votes have a signature: a tight cluster with a thin tail on one side only, sitting suspiciously close to a figure that was visible in the room. Unanchored estimates on genuinely uncertain questions are usually wider and messier than facilitators expect.
The second signature is speed. When two-thirds of a group submits an identical value within fifteen seconds of the question opening, they are not converging on truth. They are reporting the last number they saw. Response timings are one of the few honest tells you get in a live poll.
The third is what happens at re-vote. Healthy convergence moves the extremes towards the middle and leaves a visible spread. Anchored convergence collapses everything onto one value, and it usually collapses onto whichever value the facilitator repeated most often.
- ○A figure, target or previous result was visible on screen before the vote opened
- ○The question itself contained a comparison point, a range, or an example
- ○Someone spoke a number aloud before submissions opened
- ○Participants could see the running tally while the vote was still open
- ○More than half the responses arrived within the first twenty seconds
- ○The spread at round two was narrower than round one by more than half
- ○The final number sits within ten per cent of something already written in the brief
Recovering a vote you have already anchored
You will notice mid-session, usually about four seconds after the words leave your mouth. The instinct is to apologise and carry on. Do not carry on: the vote you are about to close is contaminated and you will be defending its number in a fortnight.
The recovery that works takes about ninety seconds. Name it plainly — "I gave you a number before you answered, so this round tells us what I think, not what you think" — then change the question enough that people cannot simply re-file their previous answer. Switch from a point estimate to a range, or from "how long" to "what would have to be true for this to take half that long".
Then re-run it silently and treat the first result as discarded rather than as a baseline. Keeping it on screen while the second round runs reintroduces exactly the anchor you were removing, which is the most common way a recovery quietly fails.
If the contaminated round has already produced a decision people are attached to, the honest move is to schedule a separate ten-minute slot rather than trying to re-litigate it in the last five minutes of the session. Rushed re-votes converge on the original number because nobody wants to reopen a topic at 16:55.
Where deliberate anchors are legitimate
Not every anchor is a bias to eliminate. If the budget is genuinely £40,000, hiding that from the room in the name of independence produces a lovely unanchored plan that you then have to throw away. The distinction is between anchoring a judgement and stating a constraint.
The workable rule is to sequence them: collect the unconstrained judgement first, then introduce the constraint, then ask what changes. That gives you two useful pieces of information — what the group actually believes, and what it would trade — instead of one contaminated blend of the two.
Calibration anchors work the same way. Showing what a comparable project took is valuable after a first round, where it functions as evidence people can argue with, and corrosive before one, where it functions as an instruction.
Collect the judgement first, introduce the constraint second. Do it the other way round and you never find out what your team actually thought.
Anchoring in estimates, budgets and prioritisation
Numeric votes are the worst affected because there is a continuous scale to slide along. Sizing sessions, budget splits, forecast confidence, and any question that ends in a percentage are all cases where a single early figure sets the working range for the next forty minutes.
Prioritisation votes are affected differently. In dot voting, the anchor is usually spatial rather than numeric: the item at the top left of the board, or the one the facilitator read out with most enthusiasm, collects early dots that then attract more. Running the first pass with dots hidden, or collecting allocations privately before displaying them, removes most of that.
For anything where the number will be quoted back at you in a planning meeting, run the two-round sequence. The extra six minutes is trivially cheaper than a quarter spent defending a figure that came from whoever spoke first.
Building it into how you run sessions
The reason anchoring survives in well-run organisations is that the unanchored process feels slower and less decisive in the moment. A silent round looks like dead air. A revealed spread looks like disagreement you now have to manage. Both are the process working.
Make it the default rather than a special measure for important decisions, because you rarely know in advance which vote turns out to matter. In Session Flo this means opening the poll before you discuss rather than after, keeping results hidden until submissions close, and using the second round as a genuine re-vote rather than a confirmation.
The test of whether it is working is not whether people agree faster. It is whether your first-round spreads are wider than they used to be. If they are, you are finally seeing what the room thinks.
Frequently asked questions
Does warning people about anchoring reduce it?
Barely. Anchoring persists when participants are told the anchor is random and when they are explicitly asked to disregard it, which is why the mitigation has to be structural. Change the order of the activity rather than the awareness of the room: private answers before public ones, and no visible figure until the first round has closed.
Is anonymous voting enough on its own?
Anonymity helps with the fear of disagreeing, but it does nothing about a number that was already on the screen. You can run a perfectly anonymous poll that is completely anchored. Anonymity plus hidden live results plus silent generation before discussion is the combination that works; any one of the three alone leaves a large opening.
How many rounds should a group vote have?
Two in almost every case. The first round gives you the spread, the discussion between rounds gives you the reasoning, and the second round tells you whether anything changed. A third round in a live session rarely adds information and tends to produce social convergence, where people move to end the activity rather than because they were persuaded.
Should I ever show the average of the first round?
No. Show the range and the distribution instead. An announced average becomes the strongest available anchor for the next round, and the resulting convergence looks like consensus while being a measurement artefact. If people ask for the average, give them the highest and lowest and ask what would explain the gap.
What group size makes anchoring worse?
The bias itself does not need a crowd, but its consequences scale. In a group of four, one anchored voice contaminates a quarter of the room. In a session of sixty, the first number is usually announced from the front and reaches everyone simultaneously, so you get a whole-room shift with no visible dissent. Larger rooms need the hidden-results discipline more, not less.
How is this different from the bandwagon effect?
Anchoring happens before people judge and works through a value they adjust from. The bandwagon effect happens during the vote and works through visible support for an option. The tells differ too: anchored results cluster near a specific figure, while bandwagon results show a majority that grows steeply after the first few votes land.
Keep reading
- the bandwagon effect in live polls — The sibling bias: what happens when a running tally is visible while voting is open.
- when group estimates beat experts — Why independence is the condition that makes aggregated judgement worth collecting.
- running dot voting to reach a group decision — Applies the hidden-first-pass rule to prioritisation rather than estimation.
- how consensus hides disagreement — The related failure where a group stops wanting to disagree with itself.
- Session Flo plans — Hidden live results and two-round polls are available on every plan.