WordPress AI Plugins for Personalized Content Recommendations (2026)
A machine learning recommendation engine needs behaviour to learn from, and below a certain traffic level it does not have any. A site with 5,000 monthly visitors and 200 posts gives an ML model too little signal to beat a simple tag match, which means the sophisticated option performs worse than the free one while costing money and adding a third-party dependency.
That threshold is the thing worth knowing before you shop. Rules-based related posts are free, instant, and good enough for most publishers. ML personalisation earns its place when you have enough traffic for patterns to emerge and enough catalogue for the recommendation to be non-obvious. This article covers where that line sits, prices both sides, and is direct about the privacy obligations behavioural tracking creates.
Pricing verified August 2026. Recommendation vendors frequently price on request; where a figure was not published we say so.
Rules-based versus machine learning
Rules-based recommendations match on what you already know: shared categories, shared tags, recency, popularity. They work from the first visitor, require no tracking, cost nothing, and produce results that are obvious in a good way. Their ceiling is that they cannot tell you a reader who liked article A also tends to like article F, because they only see the taxonomy.
Machine learning recommendations learn from behaviour: what people actually read next, what they buy together, what a similar visitor engaged with. They surface non-obvious pairings a taxonomy would never connect, and they need volume to do it. With thin data they default to recommending whatever is popular, which is a rules-based recommendation with extra infrastructure.
A workable rule of thumb: below roughly 20,000 monthly sessions, use rules. Above it, ML has enough to work with and the lift becomes measurable. Between those points, test rather than assume, because it depends heavily on how much of your traffic is returning visitors rather than one-off search arrivals.
The 5 options compared
1. YARPP
Yet Another Related Posts Plugin, free, and the correct starting point for essentially every publisher. It scores relatedness across titles, content, categories, and tags with a configurable threshold, which means it reads the actual text rather than only matching taxonomy, and it produces genuinely relevant results on a site with a few hundred posts.
Because it computes relatedness from your own content, it works on day one with no traffic and no tracking, which sidesteps the cold-start problem and the privacy question simultaneously. The known caution is performance: content-based matching queries can be expensive on large archives, so use the caching options, consider restricting matching to titles and taxonomy on very large sites, and check your query times after enabling it. For most sites the correct sequence is to run YARPP, measure whether readers click through, and only look at ML options if that number is disappointing rather than absent.
- Pricing: Free
- Approach: Rules-based, content and taxonomy scoring
- Best for: Almost every publisher, as the baseline
- Watch out for: Query cost on large archives; enable caching
2. Recombee
The genuine machine learning engine here, and the one to use when you have the traffic to justify it. It delivers personalised recommendations based on individual behaviour rather than content similarity, and it is designed for conversion work as much as for engagement, which makes it as relevant to WooCommerce catalogues as to article archives.
The commercial detail that changes the calculation is a genuinely generous free tier: 100,000 recommendation requests a month with almost all features available. That is enough for a mid-size publisher to run real personalisation at no cost, and it removes the usual objection that ML recommendations are an enterprise purchase. WordPress integration comes through a connector rather than a first-party plugin, so expect some setup, and understand that behavioural data leaves your site, which brings the privacy obligations covered below.
- Pricing: Free tier with 100,000 recommendation requests a month; paid above that
- Approach: Machine learning on visitor behaviour
- Best for: Sites with enough traffic for patterns to emerge
- Watch out for: Connector setup; behavioural data leaves your site
3. Contextly
Built for publishers specifically, and the most editorially minded option here. Rather than one undifferentiated block of related links, it distinguishes between related, popular, and evergreen recommendations, which is a distinction that matters enormously to a publisher with an archive worth resurfacing.
Evergreen surfacing is the feature to focus on. Most recommendation systems bias toward recent or popular content, which buries the durable pieces that would still serve a reader well three years after publication. Contextly lets you deliberately promote those, which for a site whose best work is old is worth more than marginal relevance gains. It runs on a paid subscription with a free trial and no card required; pricing is on their site rather than published in comparison articles. Judge it on editorial control rather than algorithmic sophistication.
- Pricing: Paid subscription with a free trial; confirm current pricing
- Approach: Editorial control over related, popular, and evergreen
- Best for: Publishers with an archive worth resurfacing
- Watch out for: Subscription cost against free alternatives
4. Jetpack Related Posts
Free with Jetpack, and notable for one architectural reason: the relatedness computation happens on Automattic’s servers rather than yours. For a site on modest hosting with a large archive, that removes exactly the query cost that makes content-based matching expensive locally.
The trade is control and dependency. You get very little say over how relatedness is determined, no evergreen promotion, and a feature that requires Jetpack and a WordPress.com connection. Results are decent rather than excellent. It is the right choice for a site already running Jetpack that wants related posts without adding query load, and the wrong choice if you want to influence what gets recommended or would rather not add the connection.
- Pricing: Free with Jetpack
- Approach: Rules-based, computed off-site
- Best for: Jetpack sites wanting zero added query load
- Watch out for: Little control; requires the Jetpack connection
5. Content recommendation ad networks
The widget at the bottom of news articles showing content from other sites, paid for by whoever placed it. Included because publishers routinely conflate these with recommendation plugins, and they are a fundamentally different proposition: an advertising product that happens to look like recommendations.
They pay you rather than costing you, which is the appeal, and what you give up is the visitor. Every click leaves your site, and the quality of what appears in those units is frequently poor enough to affect how readers perceive your publication. For a site monetising through advertising at scale that trade can be rational. For a site whose goal is keeping readers moving through its own archive, it is precisely the opposite of what a recommendation system should do. Decide which business you are in before installing one.
- Pricing: They pay you, per click
- Approach: Advertising, presented as recommendation
- Best for: Large ad-funded publishers
- Watch out for: Every click sends readers away; quality affects your brand
Comparison table
| Option | Cost | Approach | Works with no traffic | Tracking required | Best for |
|---|---|---|---|---|---|
| YARPP | Free | Rules, content scoring | Yes | No | The baseline for everyone |
| Recombee | Free to 100,000 requests | Machine learning | No | Yes | Higher-traffic sites and stores |
| Contextly | Subscription | Editorial rules | Yes | Limited | Archives worth resurfacing |
| Jetpack Related Posts | Free with Jetpack | Rules, computed off-site | Yes | No | Avoiding query load |
| Recommendation ad networks | They pay you | Advertising | Yes | Yes | Ad-funded publishers |
Privacy, and what behavioural tracking commits you to
Rules-based recommendations read your content. Machine learning recommendations read your visitors, and that difference has consequences worth understanding before you deploy one.
- Behavioural profiling generally needs consent in the EU and UK. Building a profile of an individual’s reading to personalise what they see is not strictly necessary for delivering your site, so it usually falls outside the exemption and requires opt-in.
- Name the processor in your privacy policy. If visitor behaviour is sent to a third-party recommendation service, that is a processor and it belongs in your policy with a description of what is shared.
- Design for the no-consent path. A meaningful share of visitors will decline. Your site should fall back to rules-based recommendations for them rather than showing an empty block, which means running both systems.
- Check data retention and location. Ask how long behavioural data is held and where it is processed. These are routine questions any serious vendor answers plainly.
- Be careful in sensitive niches. Health, finance, and similar topics mean reading history that is genuinely sensitive. Weigh the marginal engagement gain against holding that profile at all.
Rules-based systems avoid all of this, which is a real advantage rather than a consolation. If your ML lift is a few percent and the cost is a consent banner plus a processor agreement, the arithmetic may not favour the clever option.
Related guides
Frequently asked questions
Do related posts actually improve engagement?
Usually yes, on pages per session, and the gain comes mostly from having any relevant suggestion rather than from algorithmic sophistication. Measure click-through on the recommendation block itself before assuming a better engine is the answer.
How much traffic before ML is worth it?
Roughly 20,000 monthly sessions as a working threshold, and more if most of your traffic is one-off search arrivals rather than returning readers. Below that the model has too little behaviour to learn from and defaults to popularity.
Will recommendation widgets slow my site?
Content-scoring plugins add database queries; third-party services add external requests. Both are manageable, and both should be measured. Load recommendation blocks lazily below the fold rather than blocking the initial render.
Does this help SEO?
Indirectly and genuinely, because recommendation blocks are internal links, and internal links distribute authority and help crawlers find deeper pages. That said, editorially chosen contextual links inside the body of an article are worth considerably more than an automated block at the bottom.
Should ecommerce use the same tools?
Product recommendation is where ML earns its cost fastest, because bought-together patterns are genuinely non-obvious and each conversion is worth money. Recombee handles both content and products, which is why it appears here rather than in a purely publisher list.
How many recommendations should I show?
Three to six. More looks like a link farm and dilutes attention, and beyond about six the click-through on each drops faster than the total rises. Quality of match matters far more than quantity of options.
The verdict
Start with YARPP, free. It works from day one with no tracking and no consent obligations, and for most sites the results are good enough that the question of ML never arises. Measure click-through for a month before considering anything else.
Above roughly 20,000 monthly sessions: Recombee, whose free tier covers 100,000 recommendation requests a month with almost all features. That is genuine machine learning personalisation at no cost, provided you handle the consent and privacy policy work properly.
If your archive is your asset: Contextly, for evergreen surfacing. Most systems bias toward recent and popular, and deliberately resurfacing durable older work is worth more to a publisher than marginal relevance gains.
And know which direction you want readers going. Recommendation plugins keep them on your site; recommendation ad networks pay you to send them away. Both are legitimate businesses and they are opposite strategies, so choose deliberately rather than installing whichever appears first.
