Ways to Use AI for Affiliate Fraud Detection

AI gets a bad reputation from a lot of people for good reason, it gets used incorrectly.  But when used well it can streamline workflows, including helping to detect fraud in affiliate programs, and to help weed out low or zero value partners fast.

There are three prompts and examples below including:

• Finding parasitic affiliates
• Decreasing chargebacks, returns, and fake lead form fills
• Detecting and helping to prevent credit card and lead fraud

Before You Jump In

AI is not perfect, and “data” doesn’t always tell the truth, so everything needs to be checked before you take action.  You don’t want to accidentally destroy a relationship by removing a partner or reducing their commissions when in reality you should be increasing the commissions.

An example was when I had a website that showed up for the phrase “chewelry” in Google years ago.  That phrase had a 30%+/- conversion rate from my affiliate links, and the click to close time was very short as the intent was to shop and checkout fast.

It would look like a coupon touchpoint at the end of the sale from the merchant side, but it wasn’t.  AI would say it is low to no value if incorrect prompting was done, and the brand removing me would have lost a top-funnel and value-adding partner to a competitor.  This is why you need to check the output and not take things at face value.

The following information is just a few ways that you can begin using it to help save time, become more profitable, and show the real value of your efforts with affiliate management.

Note: Before you upload any data or information to any third party like AI, check with your legal team for what you can and cannot use.  You don’t want to upload and share partner or customer data with third parties like Claude, ChatGPT, or Perplexity if your legal team does not allow it.

Finding Parasitic Partners

Whether you want to remove or reward parasitic partners is up to you.  That is your decision as a business, and this prompt is for finding them.

Parasitic affiliate partners are partners that cannot drive a sale without the brand having traffic already in its funnel.  They intercept the company’s own customer, tag it with affiliate tracking, and then send it back for a commission.

For this one, having website data to run against the affiliate program data helps a lot.  You’ll have a baseline for branded and non-branded traffic, and the stats by week, season, etc… to compare individual affiliates, or affiliate groups too.  You can get this from analytics tools like ThoughtMetrics, Google Analytics 4, and Adobe.

The prompt needs to include signals that the affiliate is a parasite which may include:

• Short click to close times.
• High conversion rates.
• Rarely being a first touchpoint in a customer journey (this applies to review affiliates which are parasites, but may have value.)
• Traffic and sales patterns that match the company’s own traffic and sales patterns (i.e. when the company has an increase or decrease, the parasite magically climbs and falls with it.)

Your initial prompt may look like:

“Please look at the website stats for the average time a person comes to the website initially and then places an order, the average conversion rate for non-direct and non-branded traffic which should be lower than trademark and repeat customer visits, and time periods when sales and traffic levels are higher and lower.  Branded traffic examples can be SEO keywords like (insert brand name), remarketing audiences on meta including campaigns (insert campaign name), and Google Ads campaigns (insert campaign names).

Please compare this to the affiliates in our program which you can find (insert a spreadsheet, use a network connector, if the partner IDs are in the dashboard like datastudio or analytics that works, etc….) and find any partner that has:

• Higher than normal conversion rates,
• Shorter than normal click to close times,
• Traffic that climbs and lowers the same time the website rises and lowers,
• The number of sales increasing and decreasing at the same time frames as the website does.

Please flag any partner that matches at least two of these criteria with an explanation of what is likely happening.”

This initial prompt gets you a lot of data and saves you time when you’re just taking a program over.  The prompt can define a list, individual pages for each, or a spreadsheet.

You can also take it further by asking it then scan the partner’s website to ask if it has relevant content to the products being sold (especially if sku data is available), if they mention things like coupons and have dedicated pages to the brand, and if the website has actual SEO and social media traffic, or if it is a shell site and being used as a front for nefarious purposes.

Detecting Chargebacks, Returns, and Removing Fake Orders

Chargebacks are less common than they used to be when it comes to loyalty and cashback affiliates, but they still happen.  Returns and fake orders however can still happen as they’re harder to detect, especially with the newer affiliates that don’t have advanced detection systems.

Consumers will buy things so they get the incentive like cashback or points which clear after 30 or 45 days.  If the merchant or vendor has a 60 or 90 day return or cancel period, the consumer can return or cancel and have zero risk.  Others will take surveys when they aren’t the correct demographic for, or subscribe then cancel or downgrade to get an incentive if subscribing and upgrading is the task.

I’ve had situations on lead programs where the commission was high enough that the affiliate could have a lead filled out and an actual person have a conversation to fake it being authentic.  These leads don’t close because they are not real, so the channel becomes lower value overall.  Now the program looks like it does not have value, so everyone in it gets hurt.

There are two opportunities here.  You can recognize patterns and IDs that abuse your store, and if you’re an affiliate marketing agency like we are, you can try to detect and predict behaviors across clients to ban that ID for the clients as a value-added service.

The prompt can start like:

“I have three goals with this task:

1. Detect customer IDs that get cancelled or refunded orders regularly.
2. Find partners with higher than average cancellations, lower value leads meaning they don’t close, and higher returns.  Our cancellation rate is XY%, our average leads close at AB%, and returns are roughly AB%.
3. Have a list of order or conversion IDs so I can both block the customer from triggering more commissions but not the affiliate.

You can find the combination of affiliate and customer ID using the parameters in the URL with the affiliate ID following XYZ and customer being ABC in the URL.  If the customer ID is not available, look for any similarities in the URL structure that resulted in the lower quality conversions and share them with the affiliate ID.  This way the affiliate can locate the user and block them from accessing our program.”

Predicting Stolen Credit Cards and Low Value Leads

It’s no secret the affiliate networks have loosened their grip on fraud affiliates getting in.  AI has made it even easier for them to apply, get approved, and join programs in mass.  If you manage programs, you’re seeing it.

Pretending to own social media accounts, having tons of content and making sales but no actual SEO or social media traffic, most of the content is irrelevant, all caps for names and mismatched emails, etc…  You’ll need to upload the known accounts and then find what they have in common including emails, names, phone numbers or no number, locations, times of day they join or are activate, traffic patterns, etc…

This can all be randomized right now, but most of them don’t care and will keep going as there are more affiliate managers or brands that don’t care, or that expect the networks to be controlling this, and/or that are on auto approve (which is never smart.)  This prompt is one to help with affiliate audits, cleanups, and larger programs where accounts may have gotten traded or sold.

Your prompt can start with:

“My goal is to find affiliates that use stolen credit cards or provide low value leads as they’re likely trying to get lead commissions and are not valid customers.  Please find a spreadsheet with cancelled orders, the affiliate ID attached to the order, and (this is where you enter any other datapoints you have that are similar with fraud accounts and fake accounts in your program).  Please find another spreadsheet called (insert name) that has all available affiliate information for everyone approved into the program.

Please share with me:

• A list of affiliate IDs that match these same patterns so I can check them for fraudulent activity.
• The things they have in common including X, Y, and Z so I can look for these in applications and be proactive with preventing the issues.”

AI is a powerful tool for affiliate managements and can help streamline what would normally be labor intensive processes.  These are three you may want to try to save time when detecting and preventing affiliate fraud.  Each of the three have opportunities to build on the prompt to fine tune it and get better data.  The above is to help you with starting points, it does not mean you’ll get 100% accurate data as your situation and inputs may be different from what I would do in your situation.  I write the posts in general so people can tweak them to their specific needs.