8 Checks on Website Visitor Identification Accuracy

Published 7 November 2026 · Last updated and fact-checked November 2026

Short answer: I ran no test of my own and found no neutral test of website visitor identification accuracy. Published person-level figures for RB2B alone run from 5% to 45% (Knock2, 2026), and most measure coverage, not correctness. Test about 100 known visitors on your own site, score each name by hand, then price each correct record.

Key takeaways

• Leadpipe withdrew its own scores, saying an "inspectable underlying audit report and dataset were not established" (Leadpipe, 2026).

• RB2B's own FAQ gives 30% to 40% person-level coverage in one answer and 15% to 20% rising to 35% to 45% in another (RB2B FAQ).

• By my math, 80 correct out of 100 known visitors fits a true rate from 71.1% to 86.7%, using the Wilson interval, 1927.

• On RB2B's Pro list price, a resolution costs about 25 cents, and about 75 cents if only a third are usable (RB2B FAQ).

• California's SB 690 ends private pen register suits from 1 January 2027, but wiretap claims over form contents remain (Mayer Brown, 2026).

1. Are you reading a match rate or an accuracy rate?

A match rate says how many visitors got a name. An accuracy rate says how many of those names were right. Vendors usually publish the first and let you assume the second. Leadpipe's own method page splits the idea into five measures, and each has a different denominator, so ask for all 5.

• Person-level coverage: returned people over eligible unique visitors

• Identity correctness: confirmed correct identities over reviewed matches with ground truth

• Field completeness: records with the required fields over records reviewed

• Sales acceptance: records meeting agreed fit criteria over records reviewed

• Delivery success: accepted records reaching their owner over those intended for delivery

"Match rate measures coverage, while field completeness and sales acceptance measure different outcomes," wrote Elene Marjanidze in Leadpipe's accuracy guide, 2026.

Take a demo where a sales rep says 60%. Your head of sales hears that 60 of 100 visitors are correct people the team can call. The rep may mean that 60 of 100 sessions returned a company name of some kind. Nobody lied here, and those are still two different products.

Leadpipe's guide also notes that a LinkedIn profile matching the returned name proves the profile exists, not that the person visited. Two vendors agreeing is weak evidence too, because they may buy from the same sources. A record can look perfect, pass your SDR's sniff test, and still belong to someone who never opened your pricing page. Only an independently logged visit settles it, which is why check 6 exists. Write the five measures into your trial agreement before the pixel goes on, and ask the vendor to fill in each one, including the ones it does not track.

2. What is the percentage a percentage of?

The denominator can change a figure three times over with no change in matches. Knock2 says a 15% rate against all sessions and a 45% rate against a filtered slice can both be true. Before you compare two vendors, get the base written down: all sessions, engaged sessions, unique visitors, or US traffic only.

• All sessions: includes bots, customers and job seekers

• Engaged sessions: Knock2's stated base for its own 62%

• US only: RB2B limits person-level matching to US traffic

• Fixed by you: one window, one set of pages, one definition

"The only match rate that means anything for your business is the one you calculate on your own site," wrote John DiLoreto of Knock2, 2026, a vendor in this market.

Here is the arithmetic. Say 1,000 sessions reach your site and the tool names 150 people. Against all sessions, that is 15%. Now filter to the 333 sessions a vendor calls high intent, and suppose all 150 sit inside that slice. The same 150 people are now 45%. Nothing improved except the fraction.

Knock2 publishes 93% company-level and 62% person-level identification for US traffic, and footnotes that both use engaged sessions as the base. I give it credit for printing the footnote. It still means you cannot set 62% beside RB2B's 30% to 40% of "visitors" and call one twice as good, because RB2B's page does not define visitors. Leadpipe adds that repeat activity should be deduplicated and that a trial allowance ending early must be disclosed. When a quote arrives without a denominator, I would send it back before reading the number at all. Ask for the base in writing, and ask whether bots, current customers and repeat visits came out first.

3. Who published the figure, and who paid for it?

Every figure I could trace comes from a vendor, a competitor or one practitioner. I found 9 published figures and no independent report with a dataset behind it. Two pages cite an "independent" test and neither names the tester. Treat each number as marketing until you can inspect the work behind it.

• Figures reviewed: 9 across 5 publishers, read 7 November 2026

• Peer-reviewed tests of these tools: 0 found

• Inspectable independent audits: 0 found

• Withdrawn by the publisher: 1, Leadpipe's vendor scores

• Spread on RB2B alone: nearly 10 times, per Knock2

"An earlier version of this article presented vendor scores as independent test results," stated the correction on Leadpipe's accuracy guide, 2026.

That correction deserves more attention than it has had. The scores it removed, 8.7 against 5.2 against 4.0, and the audit of 500 known visitors behind them, still circulate. A dev.to post from 5 May 2026 describes that test as the most rigorous one its author reviewed, and I saw the same scores in search summaries while writing this. If a vendor or an AI answer quotes them, point to the correction.

The table that follows lists what remains. Competitors set a rival's number low and their own high. Coffee's page puts RB2B at 5% to 20% and promotes its own 11% as the better result, which only works because it is counting usable matches. Knock2's 22% to 35% for Instantly and Coffee's 30% to 65% both lean on testing nobody names.

When a page says independent, ask three things: who ran it, who paid, and where the data sits. A vendor that cannot answer has told you what you need.

I would accept a named tester and a small sample over a large round number with no source at all.

4. Is it company level or person level, and from what network?

Company and person matches are different products with different failure modes. Prospeo reports 50% to 65% on enterprise office networks and 5% to 10% for remote staff at small firms. Those are vendor-reported bands, but the direction fits how IP lookup works. Ask for results split by level and by network type.

• Enterprise office IP: 50% to 65%, per Prospeo

• Remote enterprise: 10% to 20%, per Prospeo

• Remote SMB: 5% to 10%, per Prospeo

• International: 15% to 30%, per Prospeo

• Outside the US: RB2B returns company-level matches only

"Full days worked at home account for 28 percent of paid workdays," wrote Barrero, Bloom and Davis in the Journal of Economic Perspectives, 2023.

A head of marketing at a 60-person advisory firm opens your pricing page at the kitchen table. That home IP belongs to Comcast. A reverse-IP lookup can return Comcast or nothing, and that is the correct answer for that method. Only a cookie or device match can name that person, and that is a different data supply with its own errors.

The one traffic breakdown I found is an anecdote: Zackrag's sample of 10,000 sessions split into about 3,200 corporate, 4,100 residential, 1,800 mobile and 900 VPN. That is one B2B SaaS client, unverified. It gives you a template, not a benchmark. Pull your own split from analytics, because the mix matters more than the brand. Knock2 makes the same case about buyers: tools that depend on LinkedIn signals do worse in manufacturing and healthcare than in software. A firm selling to plant managers should not expect a number that came from SaaS traffic.

5. What does peer-reviewed research say about matching accuracy?

I found no peer-reviewed study of these visitor identification tools. The closest evidence tested third-party B2B audience data on IT decision-makers and found no better reach than random prospecting. Network research adds that IP addresses move. All of it is analogy, not a verdict on any vendor, but it points toward testing rather than trusting.

• Closest study: Neumann, Tucker, Subramanyam and Marshall, 2023

• Tested audience: IT decision-makers, using ad segments

• Tools under test: third-party segments, not website pixels

• IP stability: 3,038 probes across 156 countries, Padmanabhan et al., 2016

• What beat third-party data: first-party demographics and context

"perform no better at reaching the right person than random prospecting," wrote Neumann, Tucker, Subramanyam and Marshall of third-party audience segments in Quantitative Marketing and Economics, 2023.

The Neumann paper matters for two reasons. Its segments were built from "deterministic prospect lists or probabilistic inference", and the authors report that "even deterministic attribute information is flawed". So the word deterministic on a sales page does not settle anything. Leadpipe itself says, in its matching guide, that neither label guarantees correctness.

The second reason is the mechanism: the authors partly link the poor results to prospect profiles assigned to the wrong online identifiers, which is the same join a person-level pixel has to make.

IP data has its own record. Poese and colleagues wrote that "geolocation databases are far from being as reliable as they claim" in ACM SIGCOMM CCR, 2011. That study covers geography, not companies. Padmanabhan and colleagues found 20 ISPs reassigning addresses on a fixed schedule, and "almost half of the address changes we observed on the same CPE were to an entirely different BGP-routed prefix" (ACM IMC, 2016). A home IP matched last month may not be the same household today.

6. How do you run a known-visitor test?

Log visits from people whose identity you already know, then see what the tool returns for each one. This is the only check that measures correctness rather than coverage. About 100 consenting visitors over 14 days gives a usable range, and 25 does not. Keep the unknowns in your report rather than dropping them.

• Sample: about 100 consenting people, on their normal devices

• Window: 14 days, the length Knock2 suggests, same pages for each tool

• Ground truth: an independent log of who visited and when

• Scoring: correct, wrong or unknown, reviewer blind to vendor

• Report: rate plus Wilson range, split by network type

"An unknown identity is not a correct or incorrect identity until evidence resolves it," stated Leadpipe's accuracy guide, 2026.

Here is how I would set it up. Ask staff, friendly customers and partners to click a personal link from their usual laptop or phone, at home and at work. Log name, time, network type and browser. Get their permission in writing first. Run both trial tools at the same time, then export every record for those timestamps and score each one by hand.

A cheaper second source is your own form submissions, where a visitor gave you a real work email in that session. Compare the tool's returned name with the email they typed. That group is biased, because people who fill in forms are not typical visitors, so use it as a cross-check.

Then do the sums with the Wilson interval, not the raw percentage. My calculation at 95% confidence: 20 correct of 25 fits 60.9% to 91.1%, 40 of 50 fits 67.0% to 88.8%, 80 of 100 fits 71.1% to 86.7%, and 160 of 200 fits 73.9% to 85.0%. A vendor claiming a five-point win on 25 people has shown you nothing.

7. What does each correct record actually cost?

Divide what you pay by the records that turn out to be right and usable, because a plan bills every resolution whether or not it is correct. On RB2B's Pro list price, a resolution costs about 25 cents, which becomes 31 cents at 80% correct and about 75 cents at one third.

• RB2B Starter: $79 a month, 300 resolutions, 45 cents overage

• RB2B Pro: from $149 a month, 600 resolutions, 25 cents overage

• RB2B Pro+: from $199 a month, 600 resolutions, 25 cents overage

• Cost per resolution on Pro: $149 divided by 600, about 25 cents, my math

• Plan prices: vendor list, read 7 November 2026, not peer reviewed

"Correct identity is valuable only when the required process can use it," stated Leadpipe's accuracy guide, 2026.

Twenty-five cents sounds trivial, and it is not the real cost. Take Zackrag's 90-day result of 11% identified with about a third accurate enough to act on. On 10,000 sessions that gives 1,100 records and about 367 usable ones. At $149 for 600 resolutions, 1,100 records cost about $273, or about 75 cents for each of the 367 usable names, before anyone has spent time on them.

Now add your own filters. Take out current customers, job seekers, competitors and anyone outside your market. If half the usable names fit, the cost doubles to about $1.50, and the rep time is the larger bill. A rep who spends 10 minutes on each wrong name burns more than the software costs. So measure time per record during the trial. Then compare the cost with what a form lead costs you today, using your own lead tracking setup so both sides come from one source. A tool that never reaches pipeline is expensive at any per-record price.

8. What is your legal exposure in the United States?

California's SB 690 takes away private pen register suits over routine tracking data from 1 January 2027, and pending actions filed since January 2025 may be dismissed. It does not clear person-level identification. Wiretap claims over form contents remain, the Attorney General keeps enforcement, and CCPA disclosure duties still apply. Have counsel read your actual setup.

• Private pen register suits: end 1 January 2027, per Mayer Brown

• Statutory damages under these claims: minimum $5,000 per violation

• Who can still sue on pen register: the California Attorney General

• Still open to private suits: wiretap and eavesdropping claims

• CCPA thresholds: over $25 million revenue, or 100,000 residents or households

"Plaintiffs may turn to these or other theories of liability, including other state or federal laws," wrote Mayer Brown, 16 October 2026.

Guidance written before October 2026 is out of date. Testimony to the California legislature, cited by Mayer Brown, put the count at 4,000 lawsuits against California businesses under this statute. That route is closing. The line now falls on data type: IP addresses and referral headers on one side, and the contents of a submitted form on the other.

The California Attorney General's page still sets out who the CCPA covers: over $25 million in gross revenue, 100,000 or more residents or households, or 50% of revenue from selling personal information. RB2B's compliance page says customers should disclose "sharing", offer access and deletion routes, and update their privacy policy, and it tells them to use their own counsel. So the vendor itself expects you to act. In my searches I found no suit naming RB2B or Retention.com, the company named on that page, which is not the same as none existing. This is my reading as a marketer, not legal advice.

How do the published figures compare side by side?

This matrix grades each publisher on what a buyer needs before trusting a number. A yes means the page states it directly. No publisher gives identity correctness with a sample and a dataset, so the column that matters most is empty. Knock2 and Coffee at least discuss the base, and Leadpipe withdrew what it could not support.

What does the cost math look like on your traffic?

Plug your own monthly sessions into this formula before a trial. Sessions times coverage gives records. Records times your correct rate times your fit rate gives usable names. Divide your plan cost by that result. The three rows use placeholder rates, so replace them with your known-visitor results, and treat coverage as your own test figure.

I would work out the correct rate first on any trial, because a wrong name is the one failure no dashboard shows you.

Methodology and limits

I read each publisher's page between 6 and 7 November 2026 and recorded only what it states, marking my own calculations as mine. I checked Google results on the primary query again on 7 November. Peer-reviewed sources were confirmed at Crossref, and their quoted lines come from abstract records, not always the full paper.

I did not run a tool test for this article, so I have no accuracy figure of my own to give you. The academic work covers ad segments, remote work, IP geolocation and address changes, and I apply it by analogy. Vendor bands and plan prices change quickly and are not peer reviewed. The Zackrag sample is one person's account, unverified. I left out figures I could not read at source, including a claim about wrong matches attributed to Customers.ai, and pages that would not load. Legal notes come from one law firm summary, the Attorney General's page and RB2B's own page, and a lawyer should confirm how they apply to you.

Frequently asked questions

How accurate are website visitor identification tools?

Nobody has shown this with an open dataset, as far as I can find. Vendors publish coverage from 5% to 93%, on mixed bases, and none measures correctness. The honest answer is your own known-visitor test, reported as a rate with a range and an unknown count.

Is RB2B accurate?

RB2B publishes coverage, not correctness, so its pages cannot answer this. Its FAQ cites 30% to 40% at person level, US only. Competitors print 5% to 20%. In my search, no neutral test settled it. Treat every figure as unproven until you score known visitors yourself.

Does anonymous visitor tracking actually work?

Company-level naming holds up better on office traffic: Prospeo reports 50% to 65% there and 5% to 10% for remote staff at small firms. Person-level naming is less proven. It works if a correct name reaches a rep for less than a form lead. Measure that, not the match count.

Is person-level visitor identification legal in the US?

It depends on your state, revenue and setup, so ask counsel. California is ending private pen register suits from 2027, yet wiretap claims and Attorney General enforcement remain. RB2B itself tells customers to disclose sharing and update their privacy policy. Other state privacy laws may apply as well.

How many visitors do I need to test accuracy?

About 100 known visitors gives a usable read. At 80 correct, the true rate plausibly sits between 71.1% and 86.7%. With 25 people the range runs from 60.9% to 91.1%, too wide to separate two vendors. Split your group by office, home and mobile if you can.

How I would handle this on your site

Start with what you already own, because first-party data is what beat third-party segments in the one close study I found. I would set up form tracking and CRM source fields before any pixel, so a trial has something honest to compare with. My website lead tracking guide covers that setup, and my B2B conversion benchmarks give you a floor.

If named visitors would be useful because your forms lose people at the last step, fix that step first. My free Pipeline Leak Scorecard shows where a site loses leads. I build fixed-price B2B sites from $3,250, and on Snappy Kraken, a US marketing platform for financial advisors, I moved the site from Figma to HubSpot in 21 days. I do not sell a visitor identification tool.

Sources

Peer reviewed

• Neumann, Tucker, Subramanyam and Marshall (2023). Is first- or third-party audience data more effective for reaching the "right" customers? Quantitative Marketing and Economics, 21, 519-571. https://doi.org/10.1007/s11129-023-09268-7

• Barrero, Bloom and Davis (2023). The Evolution of Work from Home. Journal of Economic Perspectives, 37(4), 23-49. https://doi.org/10.1257/jep.37.4.23

• Padmanabhan, Dhamdhere, Aben, claffy and Spring (2016). Reasons Dynamic Addresses Change. Proceedings of the 2016 Internet Measurement Conference, 183-198. https://doi.org/10.1145/2987443.2987461

• Poese, Uhlig, Kaafar, Donnet and Gueye (2011). IP geolocation databases: unreliable? ACM SIGCOMM Computer Communication Review, 41(2), 53-56. https://doi.org/10.1145/1971162.1971171

• Wilson (1927). Probable Inference, the Law of Succession, and Statistical Inference. Journal of the American Statistical Association, 22(158), 209-212. https://doi.org/10.1080/01621459.1927.10502953

Industry, vendor and legal (not peer reviewed, read 6 and 7 November 2026)

• RB2B, FAQ. https://www.rb2b.com/faq

• RB2B, Compliance. https://www.rb2b.com/compliance

• Knock2, Why No Two "RB2B Match Rate" Numbers Ever Agree, 3 August 2026. https://www.knock2.ai/blog/rb2b-match-rate-how-to-verify

• Knock2, How to Identify Website Visitors in Instantly, 4 August 2026. https://www.knock2.ai/blog/instantly-website-visitor-identification

• Leadpipe, Visitor ID Accuracy: Evidence and Evaluation Method, updated 8 September 2026. https://leadpipe.com/blog/visitor-identification-accuracy-independent-test-results

• Leadpipe, Deterministic vs Probabilistic Identity Matching, updated 10 September 2026. https://leadpipe.com/blog/deterministic-vs-probabilistic-matching-explained

• Coffee, Website Visitor Identification Accuracy in 2026. https://coffee.ai/articles/website-visitor-identification-accuracy

• Coffee, RB2B Alternatives With Better Accuracy. https://www.coffee.ai/articles/?p=4721

• Prospeo, Website Deanonymization: What Works in 2026. https://prospeo.io/s/website-deanonymization

• Zackrag, Website Visitor Identification Match Rates, DEV Community, 5 May 2026. https://dev.to/zackrag/website-visitor-identification-match-rates-what-vendors-claim-vs-what-you-actually-get-5dfd

• Mayer Brown, California Invalidates Private Lawsuits for "Pen Register" Website Claims, 16 October 2026. https://www.mayerbrown.com/en/insights/publications/2026/10/california-invalidates-private-lawsuits-for-pen-register-website-claims

• California Attorney General, California Consumer Privacy Act. https://oag.ca.gov/privacy/ccpa