Social proof and trust: the psychology of why we follow the crowd

When people are unsure, they copy other people

Social proof is the habit of looking at what other people did before deciding what to do yourself. It is not a weakness. It is a shortcut, and most of the time it is a sensible one. If a restaurant is full and the one next door is empty, the full one is probably better. Nobody has time to test every restaurant in town.

The shortcut is strong enough to override what people can see with their own eyes. In the 1950s the psychologist Solomon Asch sat volunteers in a room with a group of actors and asked everyone which of three lines matched a fourth in length.¹ The answer was obvious. The actors, on cue, gave the same wrong answer, one after another. Across his trials, 76% of the real participants went along with the wrong answer at least once.² Research on the experiment also shows that the presence of just one other dissenter made people less likely to conform, which matters later when we get to negative reviews.

Robert Cialdini, who spent a career studying persuasion, puts it plainly: especially when they are uncertain, people look to the actions of others to decide their own.³ Online shoppers are uncertain almost all the time. They cannot touch the product, they have never met you, and they cannot tell a solid business from a well dressed website. So they look for signs that other people chose this before them and that it went well.

Worked example: the towel card

The best known demonstration is small and practical. Hotels often leave a card asking guests to reuse towels for the sake of the environment. Researchers working with a hotel chain swapped that card for one saying that most guests reuse their towels. Participation rose from 35.1% to 44.1%.⁴ A card saying that most guests who had stayed in that same room reused their towels did better still. The researchers call this a provincial norm: the closer the crowd is to your own situation, the harder it pulls.⁵

Here is the step by step for a shop. A candle shop in Galway, Bohermore Candles (bohermorecandles.example, invented for this article), wants more customers to add a refill to their first order.

1. The generic message is "Refills make your candle last longer." That is the environmental card. It argues.
2. The crowd message is "Most of our customers order a refill with their first candle." That is the descriptive norm. It reports.
3. The provincial message, on the product page for one scent, is "Most people who bought this scent added a refill." That is the same room.

Each step brings the crowd closer to the reader. The claim has to be true, of course, which is a legal point as much as an ethical one. I come back to that below.

Six kinds of social proof, and when each one helps

Each kind of proof has a place where it shines and a place where it quietly hurts you.

The crowd: big numbers

"10,000 customers." "Over a million downloads." This is the bluntest form: sheer volume. If that many people chose it, it is probably not a disaster. It works for mass market products and for first time buyers with no other reference point.

It backfires when exclusivity is the point, when a smaller, sharper number would be more credible ("used by 40 hospital finance directors" beats "used by 100,000 finance professionals" if that is who you sell to), and when the number looks inflated next to a website that looks a week old.

Friends: people the reader already knows

"Three of your contacts use this." This is the crowd made personal. Nielsen's global survey of online respondents in 60 countries found that 83% trust recommendations from friends and family, the most trusted form of advertising it measured.⁶ Showing someone that their own network already chose you borrows that trust.

It backfires where privacy matters. Telling someone which of their contacts uses a health app or a dating site feels invasive, not reassuring. It is also weak when the connection is weak ("someone you might know") and can be a negative signal in business to business sales, where a buyer may not want to use what their competitor uses.

Experts: borrowed credibility

"Recommended by dermatologists." "Approved by a professional body." The buyer cannot judge the product themselves, so an expert they trust does it for them. This is strongest for technical, health and regulated products, where checking quality yourself is hard.

It fails when nobody has heard of the expert, when the endorsement looks bought, and when the expert is out of their field. A physicist endorsing a mattress lends nothing. It can also be unlawful when it is not true, which I cover under authority below.

Celebrities: borrowed status

A famous face lends recognition and aspiration. It works for fashion, beauty and lifestyle when the fit is real: an athlete and running shoes, yes, an athlete and accounting software, no. It fails with scandal, obvious script reading, oversaturation and an audience mismatch.

Users: people like the buyer

Reviews, ratings, testimonials and case studies. This is usually the strongest form, because the buyer identifies with other buyers more than with experts or celebrities. "Someone like me tried this and it worked" is hard to argue with. Nielsen's same survey found 66% trust consumer opinions posted online, the third most trusted format.⁶ Reviews get their own sections below, because they are where most of the research, and most of the law, sits.

Certifications: an independent check

"ISO 27001 certified." "B Corp." A certificate says an outside body checked you against a published standard. In Ireland, for instance, the National Cyber Security Centre recommends the Cyber Fundamentals framework (CyFun), originally developed in Belgium, to organisations meeting their security duties under the EU's NIS2 cybersecurity directive, with certification under it optional and a national certification system still being set up.⁷ For security, compliance and ethics conscious buyers, this is often what gets you onto the shortlist.

It fails when nobody recognises the badge, when there are so many badges the footer looks anxious, and when the certificate is out of date ("award winner 2019" invites the question of what happened since). Displaying a quality, standard or trust mark you have not been authorised to use is one of the practices Irish consumer law prohibits outright.⁸

Worked example: picking the right proof

A rota software business in Limerick, Tallyroster (tallyroster.example, invented), sells to nursing homes. It has 3,200 users, 41 nursing home groups as customers, ISO 27001 certification and a handful of glowing reviews from nursing home managers.

1. The crowd number, 3,200 users, means little to a nursing home owner. It says "popular", not "popular with people like me".
2. The sharper number, 41 nursing home groups, is the provincial norm. It goes in the headline.
3. A review from a named nursing home manager goes next to the pricing, where the doubt peaks.
4. The ISO 27001 certificate goes on the security page and in the procurement pack, where an IT person will look for it.

Same business, four kinds of proof, each placed where its reader is.

What the research says about reviews

The Medill Spiegel Research Center at Northwestern University studied how reviews change purchase behaviour.⁹ Three findings matter most for a shop owner.

First, the first few reviews do most of the work. The purchase likelihood of a product with five reviews was 270% greater than that of a product with none, and the extra benefit of more reviews diminished rapidly after those first five.⁹

Second, reviews matter more for higher priced products. Showing reviews lifted conversion by 190% for a lower priced product and by 380% for a higher priced one.⁹ The more someone stands to lose, the more they want to hear from people who went first.

Third, perfect is suspicious. Purchase likelihood typically peaked at average ratings between 4.0 and 4.7 and then fell as ratings approached 5.0.⁹ A wall of five star reviews reads as filtered, fake, or too few to include an unhappy customer.

So getting a product from zero reviews to five is worth more than getting it from 500 to 2,000.

Worked example: where reviews earn their keep

A tea shop in Kilkenny, Kieran Street Tea (kieranstreettea.example, invented), has 120 products. It has time to chase reviews for a few of them each month. Following the research:

1. Start with the products that have no reviews at all. Getting each of them to five is the steepest part of the curve.
2. Among those, start with the expensive ones: the €45 tasting set before the €6 sachet. The research found reviews matter more as the price rises.
3. Do not chase a perfect 5.0. A 4.6 average with a couple of honest three star reviews is in the range where buyers are most likely to buy.

Why a 4.8 from five reviews is weaker than it looks

Shoppers sense that five reviews are fragile, even if they never put it into words. One angry customer could move that average a long way. Statistics gives the instinct a shape. A common way to rank products fairly is a damped average, sometimes called a Bayesian average: you pretend every product starts with a few imaginary reviews at the shop wide average, then add the real ones.

$$R = \frac{C\,m + \sum_{i=1}^{n} r_i}{C + n}$$

Here $R$ is the damped rating used for ranking, $C$ is the number of imaginary reviews you add (how much weight the shop wide average gets), $m$ is the shop wide average rating, $r_i$ is the rating in the $i$th real review, $\sum_{i=1}^{n} r_i$ is the sum of all real ratings and $n$ is the number of real reviews. With few reviews, $R$ stays close to $m$. With many, the real reviews take over. It is a weighted version of the ordinary average (Foundations: the mean), and the reason a small sample swings so much is the same reason a standard error shrinks as $n$ grows (Foundations: sampling).

Worked example: two Kieran Street blends

Kieran Street Tea's shop wide average rating is 4.3. It sets $C = 20$, an assumption of the example: it treats the shop average as worth 20 reviews of evidence.

Blend A has five reviews averaging 4.8, so the ratings add up to $5 \times 4.8 = 24$.

$$R_A = \frac{20 \times 4.3 + 24}{20 + 5} = \frac{86 + 24}{25} = \frac{110}{25} = 4.40$$

Blend B has 300 reviews averaging 4.5, so the ratings add up to $300 \times 4.5 = 1{,}350$, and the same formula gives $R_B = \frac{86 + 1{,}350}{20 + 300} = \frac{1{,}436}{320} \approx 4.49$.

Blend A shows the higher raw average, 4.8 against 4.5. Ranked by the damped average, Blend B comes first, 4.49 against 4.40, because 300 people are much harder to be wrong about than five. The raw stars still go on the product page. The damped number decides the order in "best rated" lists, so a product with two lucky reviews cannot jump the queue.

def damped_rating(ratings, shop_mean, weight=20):
    """Rank products by a damped (Bayesian) average rating.

    ratings: list of star ratings for one product
    shop_mean: average rating across the whole shop
    weight: how many imaginary reviews at shop_mean to add
    """
    return (weight * shop_mean + sum(ratings)) / (weight + len(ratings))

blend_a = [5, 5, 5, 4, 5]          # five reviews, mean 4.8
blend_b = [5] * 150 + [4] * 150    # 300 reviews, mean 4.5

print(round(damped_rating(blend_a, 4.3), 2))  # 4.4
print(round(damped_rating(blend_b, 4.3), 2))  # 4.49

Negative reviews, and replies to them, earn trust

The Spiegel researchers put it bluntly in their recommendations: embrace negative reviews, because they establish credibility and authenticity.⁹ A page with a few honest complaints reads as a page nobody has scrubbed. Buyers can then decide for themselves which complaints matter to them and which do not. The Asch research fits: one dissenting voice makes people more willing to trust their own judgement.²

Replying helps too, carefully. A study of hotel reviews found that hotels which started responding to reviews saw ratings rise by 0.12 stars and review volume rise by 12%.¹⁰ Interestingly, those hotels received fewer but longer negative reviews afterwards. The authors' explanation is that unhappy customers became less likely to leave short, unfair reviews once they knew the business read them.

The reply that works is calm, specific and takes ownership. The reply that fails argues with the customer in public. Every future buyer reads that exchange, and they read it as a preview of how you will treat them.

Worked example: one bad review, two replies

A cycle repair shop in Waterford, Ferrybank Cycles (ferrybankcycles.example, invented), gets a two star review: "Service took four days, not the two promised."

1. The defensive reply: "We were very busy that week and the customer did not mention it was urgent." It is accurate, and it loses every reader.
2. The useful reply: "You're right, we promised two days and took four, sorry. A part arrived late and we should have rung you the moment we knew. We now text an update whenever a job slips." It admits the fact, gives the reason without hiding behind it, and shows what changed.

The second reply turns a complaint into proof that the shop listens. The review stays up. Hiding it would be worse, for the reason in the next section.

The Irish law on fake reviews, since November 2022

Everything above works only while reviews are honest, and in Ireland honesty is now written into the law. The Consumer Rights Act 2022 added reviews to the list of commercial practices that section 55 of the Consumer Protection Act 2007 prohibits outright, with effect from 29 November 2022.⁸ A trader must not:

• state that reviews of a product come from consumers who actually used or bought it, without taking reasonable and proportionate steps to check that they do,
• submit false consumer reviews or endorsements, or commission someone else to submit them, to promote products,
• misrepresent consumer reviews or social endorsements to promote products.

The Consumer Rights Act 2022 also made one more thing material information, which a trader must not leave out: where a business gives access to consumer reviews, whether and how it makes sure they come from people who actually used or bought the product.¹¹ The Competition and Consumer Protection Commission (CCPC) puts it plainly for advertisers: you cannot promote consumer reviews about your business unless you have taken reasonable steps to ensure they are genuine, and you must not present paid for reviews as independent.¹²

What "misrepresent" covers reads like a list of old marketing habits. The European Commission's guidance on the EU directive behind these rules names publishing only positive reviews and deleting the negative ones, giving customers pre filled positive review templates, pressing customers during moderation to change or withdraw negative reviews, and showing combined ratings worked out on undisclosed or opaque criteria.¹³ The rules exist so that consumer reviews reflect real consumers' opinions and experience.¹³

Britain's Competition and Markets Authority (CMA) has published a short guide for businesses on its own rules.¹⁴ It is not Irish law, but it makes a sensible working routine for an Irish shop as well: a published policy that prohibits fake reviews and explains how incentivised reviews are handled, a risk assessment, and processes to detect, investigate and act on fake reviews.¹⁴ In Ireland the CCPC enforces the rules, and a prohibited commercial practice is a criminal offence: on indictment, a fine of up to €60,000 for a first conviction and up to €100,000 for a later one.¹⁵

The flow reads top to bottom. The first two boxes are the paperwork the British guide expects before any review goes live, and in Ireland the policy does a legal job too, because how you check reviews is material information.¹¹ The middle is the routine for each review. The point to take away is the fork: genuine reviews go up whether they are good or bad, and only the fake or undisclosed ones come down. Hiding a genuine bad review is not a moderation choice any more. Publishing only the positive reviews is exactly the misrepresentation the ban covers.¹³

Worked example: a review drive at Bohermore Candles

Bohermore Candles wants more reviews before Christmas and plans to email last year's buyers with "Leave a five star review and get 15% off."

1. Rewarding only five star reviews is the problem. It produces reviews whose content was bought, and buyers reading them are not told. The plan changes to a reward for any honest review, good or bad.
2. Reviews written in exchange for the discount get a visible label, "received a discount for reviewing", so the incentive is apparent.
3. The shop stops its old habit of pinning the best reviews at the top. Reviews show newest first, with an option to sort by rating.
4. It publishes a short review policy page and keeps a note of how it checks reviews (only verified buyers can review, and unusual bursts get a look).

None of this costs much, and all of it makes the reviews more believable, which was the point of collecting them.

Authority: credentials, logos and the truth behind them

People follow credible, knowledgeable experts. Cialdini's examples are wonderfully mundane: physiotherapists got more patients to stick to their exercise programmes by putting their diplomas on the wall, and estate agents won more appraisals and contracts once reception staff mentioned the agents' credentials before putting a caller through.³ Credentials compress a long story ("can I trust this person to know what they are doing") into one signal.

Online, authority shows up as credentials ("chartered", "registered"), named experts and the "as seen in" bar of media logos. The logos work because the outlets have editorial standards, and being covered implies you passed them. Most visitors never click through.

That is also why faking it is tempting, and why it is dangerous. Irish law prohibits outright claiming that a business, a product or the way it trades has an approval, authorisation or endorsement it does not have, or claiming one while not complying with its terms.⁸ A logo bar of outlets that never wrote about you, or a paid advertorial dressed as coverage, invites exactly that question, and the trust damage if a customer spots it is worse than having no logos at all.

Design is the quiet form of authority. A polished site signals a competent business before anyone reads a word, which is unfair to good businesses with tired websites, and true all the same. I go into the speed of those first impressions in a separate article.

Worked example: an honest logo bar

Ferrybank Cycles was mentioned once in the Waterford local paper and once on a regional radio show. It also pays for a listing in a cycling directory.

1. The paper and the radio show go in an "as featured in" line, each linking to the actual piece. Real, checkable, local, which suits a local shop.
2. The directory listing does not go there. It was bought, not earned, and presenting it as coverage would mislead.
3. The mechanic's qualification goes on the booking page, next to the price, where a customer decides whether to hand over a €2,000 bike.

Trust signals at the checkout

Trust signals are the small things that lower the risk of handing over money: a clear returns policy, visible contact details with a real address, recognisable payment logos, a guarantee, real photos of the real people. Their absence is noticed more than their presence.

The checkout is where trust is tested hardest. Baymard Institute, which has researched checkout usability for years, puts the average documented cart abandonment rate at 70.22% across 50 studies.¹⁶ When it asked shoppers who had abandoned an order why, leaving aside the many who were only browsing, 19% said they did not trust the site with their card details.¹⁶

Each bar is one reason, and its height is the share of abandoning shoppers who named it (they could name more than one, so the bars add up to more than 100%). Two of the eight are about trust directly: the card details and the returns policy. Several others, such as surprise costs and a total that only appears at the end, feel like a trick even when they are not. A checkout that is clear about money is a trust signal in itself.

More badges are not automatically better. If you need 15 badges to convince me you are trustworthy, I start to wonder why. Placement is a better lever than quantity: a returns line and a contact number next to the pay button work at the moment of doubt, while the same items in the footer are seen by almost nobody. That is a hypothesis worth testing on your own checkout, not a law of nature.

Worked example: Ferrybank Cycles' parts shop

Ferrybank Cycles also sells parts online. It has 10,000 checkout starts a month and an average order value of €75. Today 28% of checkout starts become orders, so 72% abandon, close to the Baymard average.

It adds three things beside the pay button: the returns policy in one sentence, its phone number and workshop address, and the card logos it accepts. Suppose, as an assumption of the example, completion rises from 28% to 30%.

1. Orders before: $10{,}000 \times 0.28 = 2{,}800$.
2. Orders after: $10{,}000 \times 0.30 = 3{,}000$.
3. Extra orders: $3{,}000 - 2{,}800 = 200$ a month.
4. Extra revenue: $200 \times €75 = €15{,}000$ a month, about €180,000 a year.

Two percentage points sounds tiny. On a busy checkout it is a lot of bikes. Whether the change really caused the rise is a separate question, answered by a test, which is the last section.

Liking: people buy from people they like

We prefer to say yes to people we like, and we like people who are similar to us, who pay us genuine compliments, and who work with us towards a shared goal.³ Cialdini describes negotiation studies between business school students: in the group told to get straight down to business, around 55% reached an agreement. In the group told first to exchange some personal information and find something they had in common, 90% did, with outcomes worth more to both sides.³

For a business, liking comes from an honestly told founder story, behind the scenes content, real photos of real people and values that are lived rather than printed. It works best where the buyer is choosing a person as much as a product: services, consulting, small brands with a face. It backfires with manufactured personality, a founder in every single post, or a controversy that drags the brand down with the face. Large procurement committees care more about the contract than the founder's story.

An example: Kieran Street Tea's about page used to say "We are passionate about quality tea." It now tells the story of the owner, who started blending in a Kilkenny flat after getting fed up with stale supermarket tea, names the two farms she buys from, and shows her packing orders on a Tuesday. Nothing in it is a claim. All of it is something a customer can like.

Reciprocity: give something worth having first

People feel obliged to give back what they have received. One of Cialdini's best known illustrations comes from restaurants: a single mint with the bill raised tips by around 3%, two mints by around 14%, and one mint followed by the waiter turning back with "for you nice people, here's an extra mint" by around 23%.³ What mattered was not only the gift, but that it felt personal and unexpected.

Online, the gift is a useful guide, a free sample, a free trial or a generous free tier, and it must actually be worth having. A "free" ebook that turns out to be a sales pitch creates resentment, not obligation. A large ask straight after a small gift (a free pen, then a €10,000 contract) fails too, and if the customer feels handled, the whole thing reverses.

Worked example: the tasting sachet

Kieran Street Tea tests a free tasting sachet in first orders. Each sachet costs €0.40. It sends 1,000 first orders with the sachet and 1,000 without, and measures how many customers order again within 90 days.

1. Without the sachet, 180 of 1,000 reorder: 18%.
2. With the sachet, 210 of 1,000 reorder: 21%.
3. Extra repeat orders: $210 - 180 = 30$.
4. At an average repeat order of €24, that is $30 \times €24 = €720$ of extra revenue.
5. Assuming, for the example, a 50% gross margin, the extra profit is $€720 \times 0.5 = €360$.
6. The sachets cost $1{,}000 \times €0.40 = €400$.

The sachet worked: repeat orders rose by $\frac{21 - 18}{18} \approx 16.7\%$. It still lost €40. Reciprocity is real, and it still has to pay for itself. A smaller sachet, or one only for orders over €30, might change the answer.

Consensus: "most popular" labels and defaults

When people cannot tell which option is best, they lean towards the one most others chose. A field experiment in restaurants found that showing diners a list of the five most popular dishes raised demand for those dishes by 13% to 20%, with the effect stronger among infrequent customers, the ones with the least information of their own.¹⁷ A study of hardcover fiction found that appearing on the New York Times bestseller list led to a modest sales increase for the average book and a much bigger one for debut authors, again the case where buyers know least.¹⁸

Defaults work partly the same way. When a US employer made pension saving automatic, participation rose sharply, and many employees kept the default contribution and fund. The researchers attribute that partly to inertia and partly to employees reading the default as a recommendation.¹⁹ A preselected option says "this is what people like you normally choose". I cover defaults in more depth elsewhere.

Consensus works best under uncertainty, low expertise, risk aversion and time pressure. It works badly when exclusivity is the value, with expert buyers who want the best rather than the most common, and with customers who actively avoid the mainstream.

Worked example: a badge that worked and earned nothing

Tallyroster has three plans: Basic at €29, Pro at €79 and Enterprise at €199 a month. It tests a "most popular" badge on Pro. Each group has 1,000 new customers.

Plan	Price	Without badge	With badge
Basic	€29	400	280
Pro	€79	350	520
Enterprise	€199	250	200

1. Monthly revenue without the badge: $400 \times 29 + 350 \times 79 + 250 \times 199 = 11{,}600 + 27{,}650 + 49{,}750$, which is €89,000.
2. Monthly revenue with the badge: $280 \times 29 + 520 \times 79 + 200 \times 199 = 8{,}120 + 41{,}080 + 39{,}800$, which is also €89,000.

The badge moved people, a lot: Pro went from 35% to 52% of signups, and a chi square test on these counts gives a p value far below 0.001, so the shift is no accident. And revenue did not move by a single euro, because the badge pulled people down from Enterprise as well as up from Basic. "Did behaviour change?" and "did the business gain?" are different questions. The pricing psychology article has more on how plan layouts steer choices.

import numpy as np
from scipy.stats import chi2_contingency

prices = np.array([29, 79, 199])           # Basic, Pro, Enterprise
without_badge = np.array([400, 350, 250])
with_badge = np.array([280, 520, 200])

chi2, p_value, dof, expected = chi2_contingency([without_badge, with_badge])
print(f"chi square {chi2:.1f}, p value {p_value:.2g}")

print("revenue without badge:", (without_badge * prices).sum())  # 89000
print("revenue with badge:   ", (with_badge * prices).sum())     # 89000

Measuring social proof with one honest test

The research describes averages across many businesses, and your customers are not an average. The clean way to find out whether a piece of proof works for you is an A/B test: show half your visitors the page with it and half without, at the same time, and compare.

The headline number is the relative lift, the percentage change in conversion rate (Foundations: percentages).

$$L = \frac{p_B - p_A}{p_A}$$

Here $L$ is the relative lift, $p_A$ is the conversion rate of the original page (A) and $p_B$ is the conversion rate of the page with the change (B). A lift from 2.0% to 2.4% is 0.4 percentage points, and a relative lift of 20%.

A lift on its own can be luck. The two proportion z test asks how surprising the difference would be if the change did nothing at all (Foundations: hypothesis tests).

$$z = \frac{p_B - p_A}{\sqrt{p\,(1 - p)\left(\frac{1}{n_A} + \frac{1}{n_B}\right)}}$$

Here $z$ is the test statistic, $p_A$ and $p_B$ are the two conversion rates, $n_A$ and $n_B$ are the numbers of visitors in each group, and $p$ is the pooled conversion rate, all orders from both groups divided by all visitors from both groups. The larger $z$ is, the less likely the difference is chance. At the usual 5% significance level, a $z$ above 1.96 (or below $-1.96$) counts as significant.

Worked example: showing reviews at Bohermore Candles

Bohermore Candles tests showing star ratings and the latest reviews on its product pages. Over the same four weeks, 12,000 visitors see the old page and 12,000 see the new one.

1. Old page: 240 orders, so $p_A = \frac{240}{12{,}000} = 0.020$, or 2.0%.
2. New page: 288 orders, so $p_B = \frac{288}{12{,}000} = 0.024$, or 2.4%.
3. Relative lift: $L = \frac{0.024 - 0.020}{0.020} = 0.20$, a 20% lift.
4. Pooled rate: $p = \frac{240 + 288}{24{,}000} = \frac{528}{24{,}000} = 0.022$.
5. The bottom of the fraction: $\sqrt{0.022 \times 0.978 \times \left(\frac{1}{12{,}000} + \frac{1}{12{,}000}\right)} = \sqrt{0.021516 \times 0.000167} \approx 0.00189$.
6. The test statistic: $z = \frac{0.004}{0.00189} \approx 2.11$.
7. A $z$ of 2.11 gives a two sided p value of about 0.035, below 0.05.

So the reviews very probably helped, and the result is worth acting on. Two cautions. The test ran for a fixed period decided in advance, not until the numbers looked good. And with a lift this size, a smaller sample could easily have missed it or exaggerated it. The article on statistical significance and error margins covers both traps.

from statsmodels.stats.proportion import proportions_ztest

orders = [288, 240]        # new page, old page
visitors = [12000, 12000]

z, p_value = proportions_ztest(orders, visitors)
lift = (orders[0] / visitors[0]) / (orders[1] / visitors[1]) - 1

print(f"lift {lift:.1%}, z {z:.2f}, p value {p_value:.3f}")
# lift 20.0%, z 2.11, p value 0.035

Trust is built, not claimed

"Trust us" has never persuaded anyone. Trust is the sum of small signals, each one lowering the risk of doing business with you: the review count, the honest three star review with a calm reply, the returns line next to the pay button, the real address, the founder who actually packs the boxes.

Match the proof to the reader: reviews for considered purchases, experts for technical ones, certifications for the procurement pack, friends for social products. Aim for real and plentiful reviews, not perfect ones. Put the reassurance at the moment of doubt. Give something worth having before you ask. Use "most popular" where people are unsure, and check it earns money, not just clicks. And in Ireland, where fake and misrepresented reviews are now prohibited by name, keep it all honest because the law says so, not only because buyers can tell.

I have been optimising social proof and trust systems for over a decade, across eCommerce, SaaS and business to business work. The code in this article is a starting point for testing your own pages, not a substitute for testing them.

Footnotes

1. Solomon E. Asch, Opinions and Social Pressure, Scientific American, vol. 193, no. 5, November 1955, https://www.scientificamerican.com/article/opinions-and-social-pressure/

2. OpenStax, Psychology 2e, section 12.4 Conformity, Compliance, and Obedience, Rice University, 2020, https://openstax.org/books/psychology-2e/pages/12-4-conformity-compliance-and-obedience

3. Robert Cialdini, The 7 Principles of Persuasion, Influence at Work, https://www.influenceatwork.com/7-principles-of-persuasion/

4. University of Chicago Press Journals, A Room With A Viewpoint: Conservation Messages And Motivation, ScienceDaily, 2008, summarising Goldstein, Cialdini and Griskevicius, Journal of Consumer Research, vol. 35, no. 3, https://www.sciencedaily.com/releases/2008/08/080822160338.htm

5. Noah J. Goldstein, Robert B. Cialdini and Vladas Griskevicius, A Room with a Viewpoint: Using Social Norms to Motivate Environmental Conservation in Hotels, Journal of Consumer Research, 2008, abstract at University of Minnesota, https://experts.umn.edu/en/publications/a-room-with-a-viewpoint-using-social-norms-to-motivate-environmen/

6. Nielsen, Global Trust in Advertising 2015, Nielsen, 2015, https://www.nielsen.com/insights/2015/global-trust-in-advertising-2015-2/

7. National Cyber Security Centre (Ireland), Cyber Fundamentals (CyFun), NCSC, accessed November 2026, https://www.ncsc.gov.ie/CyFun/

8. Law Reform Commission, Consumer Protection Act 2007, revised, section 55 (Prohibited commercial practices), subsection (1)(a), (d), (i), (k), (za) and (zb), paragraphs (za) and (zb) inserted by the Consumer Rights Act 2022 with effect from 29 November 2022, updated to 24 November 2025, Revised Acts, https://revisedacts.lawreform.ie/eli/2007/act/19/revised/en/html

9. Medill Spiegel Research, How Online Reviews Influence Sales, Northwestern University, 2017 (republished 2025), https://spiegel.medill.northwestern.edu/how-online-reviews-influence-sales/

10. Davide Proserpio and Georgios Zervas, Online Reputation Management: Estimating the Impact of Management Responses on Consumer Reviews, Marketing Science, vol. 36, no. 5, 2017, https://econpapers.repec.org/RePEc:inm:ormksc:v:36:y:2017:i:5:p:645-665

11. Law Reform Commission, Consumer Protection Act 2007, revised, section 46 (Misleading: withholding, omitting or concealing material information), subsection (3B), inserted by the Consumer Rights Act 2022, updated to 24 November 2025, Revised Acts, https://revisedacts.lawreform.ie/eli/2007/act/19/revised/en/html

12. Competition and Consumer Protection Commission, Advertising, CCPC, accessed November 2026, https://www.ccpc.ie/information-for-businesses/selling-goods-and-services/advertising

13. European Commission, Guidance on the interpretation and application of Directive 2005/29/EC concerning unfair business to consumer commercial practices in the internal market (2021/C 526/01), section 4.2.4 User reviews, Official Journal of the European Union C 526, 29 December 2021, EUR-Lex, https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:52021XC1229%2805%29

14. Competition and Markets Authority (United Kingdom), Short guide for businesses: publishing consumer reviews and complying with consumer protection law, GOV.UK, 2025, https://www.gov.uk/government/publications/fake-reviews/short-guide-for-businesses-publishing-consumer-reviews-and-complying-with-consumer-protection-law

15. Law Reform Commission, Consumer Protection Act 2007, revised, sections 56 (Offence: prohibited commercial practices) and 79 (Fines and penalties), updated to 24 November 2025, Revised Acts, https://revisedacts.lawreform.ie/eli/2007/act/19/revised/en/html

16. Baymard Institute, Cart Abandonment Rate: 50 Statistics and Why People Abandon, Baymard Institute, 2026, https://baymard.com/lists/cart-abandonment-rate

17. Hongbin Cai, Yuyu Chen and Hanming Fang, Observational Learning: Evidence from a Randomized Natural Field Experiment, American Economic Review, vol. 99, no. 3, 2009, https://www.aeaweb.org/articles?id=10.1257/aer.99.3.864

18. Alan T. Sorensen, Bestseller Lists and Product Variety, Journal of Industrial Economics, vol. 55, no. 4, 2007, https://econpapers.repec.org/RePEc:bla:jindec:v:55:y:2007:i:4:p:715-738

19. Brigitte Madrian and Dennis F. Shea, The Power of Suggestion: Inertia in 401(k) Participation, Quarterly Journal of Economics, vol. 116, no. 4, 2001, https://econpapers.repec.org/RePEc:oup:qjecon:v:116:y:2001:i:4:p:1149-1187