top of page

What is A/B testing? Examples from the Liberal Democrats during the 2019 general election

You may have seen elements of this report in this Observer/Guardian news piece and then also reported on the Andrew Marr show. The following is a detailed report on A/B testing by the parties that those pieces were based upon.

 

Political parties using Facebook ads have a powerful capacity to hone their messages. Unlike the past where expensive and hard to organise focus groups were the only avenue available for parties to test their messages; today all the political parties engage in what marketing calls A/B testing (aka multivariate testing).


Political parties have never had it so good, as it’s ‘us’ (the users), who help make this happen. Today party campaigning is Janus-faced, a fantastic example of this phenomenon is via the many faces of Jo Swinson the Liberal Democrats were recently pushing.

Image 1 — From October 29th to 1st November 15 “Stop Brexit with Jo Swinson” adverts have been sent to a combined 128,000 people through a spend of £2,985

You may be surprised when browsing Facebook to see the same advert twice but with slight differences, you might not even notice. However, for the parties, we the Facebook audience are in fact lab rats in a giant experiment. Through your activity on Facebook, your data is helping political parties decide on the messages and visuals they will use throughout the general election campaign. The photo used on the front of a leaflet, the image used within a poster, or the messages used to campaign are no accident. They have been ground tested repeatedly to a purified form – using us as the catalyst.


So what is A/B testing?

Image 2 — A simple example of A/B testing courtesy of Wikimedia — the only difference that leads to higher clicks is green colour and an arrow

A/B testing is a method of comparing two or more versions of an advert against each other to determine which one performs better. Depending on what content is used, people will interact differently. Then statistical analysis can be used to examine which variation performs better for a given goal. The simple example used in Image 2 belies the fact that goals can be multi-faceted. They can cover a wish to influence certain demographics such as younger people, the region reached, or how many clickthroughs can be achieved.


All the parties are engaging in this process, however clear examples have shown up recently via the Liberal Democrats. The Liberal Democrats page is currently spending a lot on Facebook, with £25,502 spent from the 27th October to the 2nd November 2019. Some of this money is being used to A/B test a variety of messages, images and policy approaches. Of this sum, £2985 was spent sending 15 different versions of Jo Swinson to a combined 128,000 people. Jo is presented from smiley and determined to thoughtful and humorous. All the posts feature the same message; “Liberal Democrats have led the fight to stop Brexit for over 3 years. We’re not giving up — and neither should you”, and a link to the Liberal Democrat website.

Image 3–15 versions of Jo Swinson

Given the increasing importance of leaders alongside the role of the personalisation of politics in campaign strategy, the Liberal Democrats are clearly trying to find the right Jo for their campaign communications. Having the best images of the leader for content is today imperative, as research consistently shows voters appreciate leaders as communicators with leaders a draw for engagement. The Liberal Democrat’s interest in A/B testing is however not just pushed by an interest in presenting the leader, but is also spurred on by several key political factors. So what are they trying to achieve and why?


Reaching the right people in the right places


Using Facebook Ad Archive data we can examine the basic demographics of those who received the 15 adverts. Graph 1 below shows how each of the adverts received different audiences, with some pronounced differences between age groups and genders. The party is generally reaching more women than men, although some of the adverts see the opposite trend. Overall, the Liberal Democrats are reaching distinct audiences using different Jo Swinson images. This is important as one key goal of the adverts is to prompt clickthroughs, with parties interested in informing, activating and organising people dependent on their political position, location and socio-demographics. A/B testing allows the parties to reach and influence who they believe to be the most important voters.


Graph 1 — Men and Women’s audience demographics for the Adverts

N.B — data does not add to 100%, this is partly due to ‘unknown’gender, alongside Facebook’s reports being inconsistent

As Graph 2 shows the party is generally reaching different numbers of women to men depending on the image used. This is because of the Facebook ad delivery algorithm, user choices and targeting parameters chosen.


Graph 2 — Advert audience difference, Women’s demographic difference to Men’s

Women are being reached more by the Liberal Democrats adverts, with this seen especially in older women

Improving their communications


In A/B testing the best performing posts (according to goals) are continued while others are shut down. Of the 15 adverts sent by the Liberal Democrats, they decided to retire 10 on the 31st of October and continue with 5.

Image 4–10 Jo Swinson images did not survive the process

These 5 adverts that were still running until November 1st received no extra money, but gained more impressions (around 3000 more impressions each than ad’s 1 to 10). As Graph 2 shows, the Liberal Democrats choose to continue using 5 adverts of Jo Swinson that were reaching a more equally gender balanced audience, featuring a slightly younger audience with more younger men than women.

Image 5 — A heavily used image of Jo Swinson seen amongst the 15 adverts

Of the remaining 5 images it is important to note that Jo 12 (see Image 5) has been used heavily recently by the party across both Twitter and Facebook (see Image 6), while Jo 15 has been seen on the Liberal Democrat campaign bus and on leaflets (Image 7). Recent use will likely be highly related to the clickthrough rates achieved via the adverts. The party is clearly trying to find images of Jo that energise the right sort of people to click, activating electorally important voters. At the moment this looks to be younger men and women, as well as older women (Graph 2). These groups are important demographic targets for the Liberal Democrats given the way the electorate voted in the 2016 EU Referendum.

Image 6 — 3 instances of Jo 12 being used in Liberal Democrat communication

Image 7 — Jo 15 as seen on leaflets via — https://electionleaflets.org

Although choosing the right advert image to campaign is a central outcome of A/B testing, other goals are also at play. A central one is gathering data.


Gathering data


Although, the 5 adverts of Jo Swinson that were kept running are potentially more alluring to an audience meaning they are more likely to use these images in communications (as shown above). A main reason the Liberal Democrats will have selected Jo 12 and 15 over the others is due to more clickthroughs. In the early stages of an election campaign a party’s true goal is often data gathering. All the 15 Jo adverts sent users through to the Liberal Democrat website shown below (Image 8). This website is clearly designed to gather user data including where they live. The Liberal Democrats can then match this data to other data sources. The website also asks users for emails, helping the party generate an emailing list.


Using different Jo Swinsons with potentially differing targeting parameters allows for the Liberal Democrats to maximise data gathering across electorally important seats and demographics. Thus, for the Liberal Democrats, the best images of Jo Swinson are those that prompt the greatest amount of data gathering.

Image 8— the Liberal Democrat website you are sent to if you clicked

So what does this all mean?


Facebook has allowed parties to engage in complex psychological experiments, hyper-charging their content to maximise their intended goals. This has both positive and negatives, it gives parties greater opportunities to activate support or reach and inform the public of their policy positions, issues and candidates. The practice may seem very innocent when content is positively charged as in the Jo Swinson example, or when testing colour choice as Image 9 below shows. However, A/B testing is not just used to promote the best images of the leader or choose the best visuals.

Image 9 — 9 adverts testing different title colours

A central approach is via negative campaigning. This describes a party using negative content to attack the opposition, it can lead to a more negative campaign environment reducing voter turnout and increasing radicalising politics. In the past, negative advertising was much harder to use. Testing messages via focus group was expensive and the chosen messages were then generally sent at the broadcast level. Broadcast approaches meant you had as much chance to put someone off voting for you as pushing someone to stop voting for your opponent, as being associated with negative campaigning could in turn isolate undecided voters and reduce trust. Today, Facebook has allowed parties a safe space to engage in negative messaging. These messages can be tested and improved to increase their effects, they can be sent to certain voter groups without the wider public ever knowing. Facebook data is thus contributing to making our politics a far more obsequious and negative affair.


As seen in Image 10, the Liberal Democrats are placing a lot of effort in finding the best line to attack Jeremy Corbyn. This works because negative posts also feed through to the same party website seen above, the Liberal Democrats can use clickthroughs to measure what are the best lines or images to attack opposition leaders and activate their support.

Image 10 — 17 examples attacking Jeremy Corbyn

Facebook data from our interactions with this negative personalised content is used by parties to find the best attack lines. This means that parties attack messages can be developed and sent to those who it will affect most. This is a truly new capacity given the ease with which certain messages can be linked to certain groups. For example, Graph 3 below shows that these negative anti-Corbyn messages are being focused on younger/middle aged men.


Graph 3 — Men and Women’s audience demographics for Corbyn attack adverts

Graph 3 — Negative Corbyn Adverts and associated audience

It is important to note that it is not just the Liberal Democrats who engage in this process, examples of positive and negative A/B testing are visible from both the Conservative Party and Labour Party. Their testing appears to have occurred earlier on from August to early October, before the election campaign properly started. It is clear that both these parties knew an election was going to happen eventually.

 

Below (Image 11) one can see a subset of advert images sent by the Conservatives from 8th to the 13th of August 2019. Across dozens of adverts different versions of Corbyn as well as floating arms, are seen ‘stuffing the ballot box’ with remain tickets. All the posts featured this text that speaks directly to a people versus Parliament narrative:

Politicians like Jeremy Corbyn are only respecting remain votes — and ignoring 17.4 million leave voters. ❌ Add your name now — don’t let him get away with it. ⬇️

Image 11 — Corbyn stuffing remain ballots into a ballot box

The colours being tested are interesting as they have also recently popped up in posters and other avenues. The Conservative Party have taken on the Vote Leave red colour, with this recently being used at the Conservative Party campaign launch. All the adverts fed through to this website (Image 12), which as seen in the Liberal Democrats example, is all about gathering data.

Image 12 — The Conservative website linked from the adverts

These adverts were again aimed at men. Why negative personalised adverts appear to be focused centrally on men is interesting and has not been a focus of campaign and communications researchers. It may be due to Brexit and the fact that men voted more heavily for leave, however given how both the Liberal Democrats and Conservatives push Corbyn focused content towards men this is an area requiring further investigation. The pushing of negative content towards men over women is fascinating and problematic. Parties appear to be driving a wedge between men and women in the content they send. Approach does not appear to be based on content, but instead to be based purely on gender and emotion.

Examining a sample of 22 of these adverts in Graph 4, all the adverts were sent for only one day. 17 of the adverts received under £100 funding, reaching around 1000–5000 people each. However a group of 5 adverts received more funds (£100–500) reaching 10,000–50,000 people each. Of the adverts that received more funds these were sent more heavily to men and less to women.


Graph 4 — Men and Women’s audience demographics for Conservative Anti-Corbyn adverts by spend

A further example of Conservative A/B testing can be seen in the below wall of posts associated with policing (Image 13). All the posts asked:

🚨More police = less crime ⬇️ That’s why we’re hiring 20,000 new police officers, starting right now. Want to see more in your area? Let Boris Johnson know right now 👇

Again dozens of examples exist with the adverts sent on the 7th August 2019 and then stopped. The adverts feed through to this website that pushes the new police officers, but again asks you to give your ‘name, email address and postcode’. The adverts did not receive much ad spend individually, most only cost under £100 and reached 1000–5000 people.

Image 13 — Conservative Party policing advert images of numerous varieties

Labour also consistently use A/B testing, although most of their examples see only 2 or 3 advert varieties tested, their best example of negative testing is seen in 3 images linking the NHS to Donald Trump (Image 14).

Image 14 — Labour Trump NHS attack adverts

These adverts attacking Trump, Jo Swinson and Boris Johnson ran from the 25th October –31st October 2019 and were generally aimed at younger women and men. Several versions of each were seen, with the adverts not receiving much ad spend with most under £100.


These posts, like the policing posts seen via the Conservatives may be testing the water for a later campaign approach. This is because these adverts are not being spent heavily on. Instead they are being A/B tested with the potential that they (or messages informed by these adverts) may be rolled out heavily later in the campaign. The adverts may however just be to do with gathering data.


Other examples via Labour are also available. As seen in Image 15, several different adverts based upon Brexit were sent across one day in late October, with the content feeding through to another party website intent on information collection. Almost all the posts seen used this text:

Brexit should not be a choice between the Tories’ bad deal and a disastrous No Deal. That’s why Labour would put any Brexit deal back to the people to let them decide. Agree with us? Add your name to our petition ✍

Image 15— Five examples of Labours advert images asking users to sign a petition asking for a referendum

As well as name, email and post code as seen via the other party websites, the website (Image 16) even asks for mobile numbers. These contact details are potentially utilised later on for mass Whatsapp messaging, texting or phone conversations. Momentum says it reached some 400,000 people through WhatsApp in 2017 so it can be an important get out the vote (GOTV) tool.

Image 16 — The website the adverts were linked to, 17,310 people signed the petition

 

Overall all the parties are engaging in A/B testing, with this split across uses for improving their messages, images and data collection. However as well as improving positive messages and policy platforms, making sure messages are reaching the right people effectively. A/B testing is also used in a negative manner to hyper charge content designed to demobilise voters. As Pippa Crerar picks up on when commenting on our Observer article: “ultimately, the adverts that seem to be getting the most resonance amongst voters are those that are attack adverts”.

This ability for elections to be fought as much on demobilisation as mobilisation is a threat to the very nature of democracy, as negative messages that can drive voters to polarising opinions are becoming more refined. We the voters are being compartmentalised, fractured apart and sold different ideas, with this having stark impacts on the foundations of community and the common ground of politics.


It is our activity online that makes this process possible, we have a right to understand how it works and what the parties are doing with our information. However, if we want to change things this must be done at the regulatory level, as the problems associated with data-driven negative advertising will likely only get worse. The tools available are powerful and yet we the public have not been made aware of what is going on, this report is a first step in opening up a discussion about what is going on and the ramifications it may have.

Comments


bottom of page