Behavioral targeting means changing what a visitor sees based on what they have already done: pages viewed, items left in a cart, past purchases, or whether they are new or returning.
On your own site it works best when the change answers a question that visitor actually has, such as a cart reminder for someone who abandoned a cart.
Test it against a control inside the same segment, because a targeted variant that "wins" without one usually just shows that the segment was already more likely to buy.
Geo targeting is the location-based version of the same idea.
It is one of the most reliable patterns in our test database, mostly because shipping, currency and delivery times really do differ by country.
What behavioral targeting is, and what it isn't
The term comes from advertising. The FTC's February 2009 staff report on online behavioral advertising describes tracking a consumer's online activities over time, including searches, pages visited and content viewed, to deliver ads matched to their interests.
The same report separates out "first party" behavioral advertising, meaning targeting by and at a single website, and contextual advertising based only on the current page or search.
FTC staff judged first-party targeting more consistent with consumer expectations and less likely to cause harm, and contextual advertising less invasive.
Most conversion work sits in that first-party zone: a store using its own data about its own visitors.
On a website, behavioral targeting usually draws on a handful of signals:
- Visit history: new vs returning, number of sessions, days since last visit.
- On-site actions: products viewed, cart contents, quote or application started but not finished.
- Purchase history: past buyer vs never purchased, category bought, subscription status.
- Entry context: traffic source, ad campaign, landing page. This overlaps with ad message match.
- Location: country, region or city, inferred from the IP address.
It is a subset of personalization. Rule-based targeting ("show X to segment Y") is different from AI personalization, where a model picks content for each visitor, and from product recommendations, which personalize a single slot.
What the tests show for behavior-based targeting
The clearest wins in our database come from reminding people of something they already started.
Visually reported that a cart reminder for returning Chêne Gear shoppers lifted conversion rate by 10.6%, and a homepage cart reminder for Revant Optics cart abandoners lifted revenue by 7.6%.
Lead-gen sites see the same pattern. A North American insurer working with AB Tasty showed returning visitors who had abandoned an application a quote widget reminding them of their quote, and applications submitted rose 23.47%.
Targeting by intent level also shows up. OptiMonk reported that BOOM! by Cindy Joseph used segmented popups with a 10% coupon for high-intent visitors and a lead magnet for everyone else, adding $148,297 in 30 days.
That was a before-after comparison, not a controlled test, and some coupon orders would likely have happened anyway.
New-visitor targeting tends to work when it explains the product.
Visually reported +2.6% conversion rate for an Underoutfit banner showing new visitors the 180-day returns policy, and TAG reported +3.4% for a 'How It Works' homepage section for Modibodi's new visitors.
The lifts are modest, and that is the honest picture.
Bigger claims exist, but they often don't hold up.
Clean Commit's homepage tailored to new vs returning visitors reported +241.1% conversion rate but did not reach significance, with overlapping tests and holiday promotions interfering. We count it as inconclusive.
VisuallyVisuallyGeo targeting
Geo targeting changes content based on where the visitor appears to be.
For ecommerce that usually means currency, shipping costs, delivery dates, duties and returns.
For lead-gen and SaaS it often means local proof: customer logos, landmarks, offices or event locations near the visitor.
How accurate is location data?
Location almost always comes from the visitor's IP address, which is a good guess, not a fact.
MaxMind, a major IP geolocation provider, estimates 99.8% accuracy at the country level.
For US IP addresses it estimates about 80% accuracy at the state level and 66% at the city level, within 50 km.
Its accuracy comparison tool shows similar numbers for its US city database: 60% of IPs placed within 50 km, 32% placed in a city more than 50 km away, and 8% with no city data at all.
VPNs, privacy networks and mobile carriers make all of this worse.
The practical rule follows from those numbers:
| Level | Reliability | Safe uses |
|---|---|---|
| Country | High (99.8% estimated) | Currency, shipping rates, delivery estimates, local logos, language defaults |
| State or region | Moderate (about 80% in the US) | Seasonal or regional messages that still make sense if slightly off |
| City | Low (66% within 50 km in the US) | Soft touches only, like an image; never prices, stock or "near you" promises |
Always give visitors a way to correct a wrong guess, and fall back to generic content when the city is unknown.
What geo targeting tests show
Country-level shipping and currency changes are the most consistent wins we have.
Dave Diederen reported +13.01% revenue per visitor from a geo-detection popup asking international visitors to confirm their country and switch currency, and +4.88% RPV from currency codes next to prices based on shopper location.
Shipping copy that names the visitor's country also does well.
Dave Diederen reported +10.42% conversion rate after changing "Free Shipping" to the country-specific 'Free Shipping To DE', and +13.48% RPV from a shipping bar with a 'Ships by' date and a localized country flag. See shipping messaging for the wider pattern.
At checkout, One Click reported a +30% conversion rate from a country-localized checkout at Readers.com, where Canadians saw Canadian shipping rates, delivery estimates and currency. No sample size or significance was given.
Outside ecommerce, Humana reported +31% clickthrough after showing a couple in front of a local landmark in its homepage banner, such as the Golden Gate Bridge for Bay Area visitors.
VWO reported +376% leads on its own free trial page and +247% on its demo page after showing country-specific customer logos to visitors from nine countries.
That is vendor-reported, with the variant changed mid-campaign for three countries, so treat the size as unproven.
Seasonal timing can be local too. Visually reported +40.8% conversion rate for Chêne Gear from a countdown to the visitor's state hunting season opening.
It works because the season opens on different dates in different states, so a single national timer would be wrong for most visitors.
Where geo targeting stops helping
The Global Leadership Summit ran two geo tests with different results.
An email naming the closest host site produced 87 more registrations than the control.
A localized landing page that asked visitors to pick their location first raised clickthrough 32% but didn't change registration rate.
The lesson: local relevance raised engagement in both tests, but only one turned into more sign-ups.
Measure the outcome you care about, not the click.
Audience targeting A/B testing
Audience targeting and A/B testing answer different questions.
Targeting decides who sees an experience. Testing decides whether that experience beats the alternative for those people.
You need both, and most mistakes come from skipping the second.
Randomize inside the segment
The right setup takes everyone who matches the audience rule, for example returning visitors with an abandoned cart, and randomly splits them into control and variant.
Both groups then share the same intent, so the difference between them measures the targeted content.
The wrong setup shows targeted content to the segment and compares it with everyone else.
Returning cart abandoners convert better than average visitors no matter what you show them, so the "lift" is mostly the segment, not the change.
Once a targeted experience wins, keep a small holdout group that still sees the default.
That lets you check months later that the effect is still there.
Segments shrink your sample
Targeting means only part of your traffic enters the test, so plan for a longer run.
Say a store gets 100,000 visitors a month and 30% are returning visitors.
A returning-visitor test receives 30,000 visitors a month, or 15,000 per arm in a 50/50 split.
If the sample size calculator says you need 45,000 per arm, that test takes three months instead of under a month.
A country-level test for a market that sends 5% of traffic may never finish.
In that case, roll out changes that are plainly correct, like the right currency, without a test.
When a targeted test is the wrong tool
- When the change is a fix, not a hypothesis. Showing Canadians prices in Canadian dollars is getting a fact right. You don't need a test to confirm that.
- When the segment is too small. If the test can't reach its sample size in a reasonable time, a test is the wrong tool. Run the change on a larger audience or skip it.
- When you slice results after the fact. Running a site-wide test and then hunting for the segment where the variant won is a multiple comparisons problem. Decide the segment before the test starts and confirm any after-the-fact finding with a new test.
- When segments overlap with other tests. A returning-visitor test running alongside a site-wide test can produce interaction effects. Clean Commit's inconclusive personalization tests list overlapping tests as one of the problems.
Mixing segments has a statistical trap too. A variant can win for new and returning visitors separately yet appear to lose overall if the traffic mix shifts between arms, which is Simpson's paradox.
Keep the split ratio fixed for the whole test.
A simple order of work
- Start with country-level fixes: currency, shipping costs, delivery dates. Roll these out directly.
- Then test reminders for visitors who left something unfinished: carts, quotes, applications.
- Then test new-visitor explanations: how it works, returns policy, proof.
- Leave fine-grained behavioral rules and city-level content for last, when you have the traffic to test them.
For how to define audiences, see segmentation. For swapping content blocks per audience, see dynamic content, and for software that does the targeting, see website personalization tools.