removal services, data brokers, and the right to be forgotten

(authored by agents unless marked 🧑)

short version

• words
  • data broker: a company that collects and sells personal data about people who are not its customers
  • people search site: a data broker whose website lets anyone look a person up by name
  • removal service: a paid service that sends removal requests to brokers for you, e.g. DeleteMe, Incogni, Optery
• what the literature already shows, 2022–2026
  • asking is hard and many brokers ignore you
  • paid removal services remove about a third to a half of what they find
  • asking for removal makes you hand over more personal data
• what nobody has measured, as far as I found
  • whether a removed profile stays removed
    • every study stops at 2–4 months
  • whether data that was supposed to be deleted shows up in later breaches
    • the human’s idea; one famous case exists, no systematic study found
  • whether California’s new one-stop deletion portal works for real people
    • brokers had to start processing its requests on 1 Aug 2026
    • the 90-day window concerns status reporting, not a deletion deadline
• my three strongest ideas, details under “research ideas”
  • 1: watch real profiles before and after a California portal request, and keep watching for a year
  • 2: use public breach notices to count how much breached data should already have been deleted
  • 3: plant made-up identities with unique email addresses to see whether opting out leaks your data
• these rankings are my opinion; no prototype or pilot exists

scope and evidence

• reviewed 7 Oct 2026 UTC, starting from the human’s note
  • 🧑 the idea I build on: “find out if exercised when data breach occur”
• evidence labels used below
  • read: I read the relevant sections of the full text
  • abstract: I read only the abstract
  • snippet: I only saw a search result summary; check before relying on it
• carried over from the Codex agent’s first version: its close reading of the Consumer Reports study
• ChatGPT consultation was not possible: the tool failed with “account_ui_login_required”
• I did not chase every citation chain; law review articles on the right to be forgotten are mostly left out

how removal works today

• people search sites build profiles from public records, other brokers, and web crawls
  • He et al. 2025, read: “web crawls, public records, online activity, self-reported information and other data brokers”
• to get out you ask each broker separately, or pay a service to ask for you
• the laws
  • EU GDPR article 17: right to erasure, also called the right to be forgotten
  • California CCPA: right to delete, and right to opt out of sale
    • a delete request needs identity verification; an opt-out must not require it
  • California Delete Act: brokers must register; the state runs one portal named DROP that sends your delete request to every registered broker
    • correction checked 8 October against CalPrivacy’s instructions: “Data brokers have up to 90 days to report how they processed your request”
    • the 90-day status-reporting window must not be described as the deletion deadline
    • official instructions separately describe continuing checks and deletion of newly matching data at least every 45 days
    • same article: “Currently over 600 data brokers are in the registry”
    • snippet: the portal opened to residents on 1 Jan 2026; “more than 290,000 deletion requests have been submitted”
• search engines are a separate path: Google can hide a result without the source page changing

what prior work found

do removal services work?

• Grauer, Kauffman, Honeywell, Data Defense: Evaluating People-Search Site Removal Services, Consumer Reports, 2024; read by the Codex agent
  • 32 volunteers from California and New York, 7 paid services plus a do-it-yourself group, 13 people search sites, checks after 1 week, 1 month, 4 months
  • paid services: 117 of 332 profiles gone after 4 months, about 35%
  • do-it-yourself: about 70% gone
    • the report’s text and its table 2 disagree slightly, 36/47 vs 70%; use the raw counts
  • limits, in the report’s words: “We also did not test for the reappearance of profiles” and “not statistically significant or nationally representative”
  • once a profile disappeared they stopped checking it, and they only checked the original URL
• He, Snyder, Haddadi, Bustamante, Tyson, Measuring the Accuracy and Effectiveness of PII Removal Services, PoPETs 2025; read
  • 10 services surveyed, covering 2,024 brokers; 71 participants used 4 of the services
  • services disagree on which brokers matter: “average Jaccard similarity of 0.21”, “only 10 data brokers common to all services”
  • most “found” records are someone else: “only 41.1% of records identified by these services were PII about themselves”
    • so a service may remove strangers’ records, and may hand your data to brokers that never had it
  • “only 48.2% of records successfully removed”
  • outcomes come from what the service’s dashboard says and what participants report; the authors did not watch broker sites themselves over time
  • they ask for “free alternatives that streamline data removals”
• snippet only, unverified vendor-style claim: “61% of records reappear within six months after opt-out”
  • I could not find a paper behind this number

do brokers obey requests?

• Take et al., What to Expect When You’re Accessing, PoPETs 2024; read
  • researchers sent access and removal requests for themselves to 20 people search sites
  • “four main groups are behind 14 of the sites studied”
    • removing from one site in a group can remove from its sister sites
  • “No reappearance was observed after two months”
    • a handful of researchers, one follow-up check
  • sites refuse access by calling the data public: the CCPA definition of personal information “does not include publicly available information”
• Take et al., “It Feels Like Whack-a-mole”, PoPETs 2022; abstract
  • interviews with 18 people who tried removal; they report that data comes back
  • reports, not measurements
• van Kempen et al., Consumer Beware! Exploring Data Brokers’ CCPA Compliance, IEEE S&P 2026; abstract
  • access requests to “all 543 officially registered data brokers”; “Above 40% failed to respond at all”
  • brokers “requested personal information as part of their identity verification process, including details they had not previously collected”
• van Kempen, Tsudik, Jhunjhunwala, Raja, Let My Data Go, arXiv Jul 2026; read
  • delete and opt-out requests to 322 registered brokers using two made-up people
  • only 97 of 322 answered the delete request with a result
  • 22% verified identity for opt-outs, which the rules do not allow
  • some answered that data “was deleted” for a person who never existed
  • their own stated limit: “it did not evaluate the actual efficacy of submitting delete and opt-out requests. Doing that would require using real identities”
  • their stated future work: evaluate DROP “once it is fully operational”, and repeat with real people
    • so this group will likely do a DROP study; anyone else has to be fast or different
• Sun, Vekaria, Nithyanand, On the Suitability of LLM-Driven Agents for Dark Pattern Audits, PoPETs 2026; abstract
  • an LLM browser agent walked through access request forms on “456 data broker websites”
  • it audits the forms for manipulative design; it does not test whether deletion happens
• Consumer Reports, authorized agent study, 2021; snippet
  • opt-outs for 124 Californians to 21 companies; 12 of 21 confirmed they stopped selling

does deletion happen at ordinary online services?

• Rupp, Syrmoudis, Grossklags, Leave No Data Behind, PoPETs 2022; read
  • made accounts at 90 services, asked for erasure, then “More than six months later” sent access requests “to find out if and what data remains”
  • “At 27%, the share of non-compliant services is not negligible”
  • limit: a service that kept the data can simply answer that it has none
• European Data Protection Board, 2025 coordinated enforcement report on the right to erasure, Feb 2026; snippet
  • 32 regulators questioned 764 organizations; recurring problems include unclear retention periods and data left in backups
  • based on what organizations say about themselves
• Potter et al., SoK: The Gap Between Data Rights Ideals and Reality; abstract
  • reviews “201 interdisciplinary empirical studies, news articles, and blog posts”; a good map of the wider field
• Bertram et al., Five Years of the Right to be Forgotten, CCS 2019; snippet
  • Google’s own data on 3.2 million URLs people asked it to hide; “1,000 requesters generated 16% of requests”

the removal process is itself a risk

• Di Martino et al., Personal Information Leakage by Abusing the GDPR ‘Right of Access’, SOUPS 2019; snippet
  • pretended to be someone else at 55 organizations; 15 handed over that person’s data
• Pavur and Knerr, GDPArrrrr, 2019; snippet
  • same attack against more than 150 businesses
• Krebs, Mozilla Says It’s Finally Done With Two-Faced Onerep, 2025; snippet
  • the founder of removal service Onerep also founded people search sites, including Nuwber
• National Public Data, a background check broker, was breached in 2024; about 2.9 billion records leaked, per news reports; snippet
  • a broker breach makes later removal from that broker pointless

what people do after a breach

• Mayer, Zou, Schaub, Aviv, “Now I’m a bit angry”, USENIX Security 2021; read
  • showed 413 people real breaches that exposed their email address, using Have I Been Pwned
  • “Participants were unaware of 74% of displayed breaches”
  • of the breach cases where the person described what they did: “18 (13%) deleted or deactivated the account”
    • the same passage gives “87, 61%” for password changes, so the base is about 143 cases
    • self-reported; says nothing on whether the company then erased anything
• Zou et al., “I’ve Got Nothing to Lose”, SOUPS 2018; read in part
  • 24 interviews after Equifax: “few knew whether they were affected, and even fewer took protective measures”
• Bhagavatula and Bauer, (How) Do people change their passwords after a breach? and What breach?; read in part
  • real browsing and password data instead of surveys
  • “only 33% of the 63 changed their passwords”
  • “only 16% of our 303 participants visited an incident-related web page”
• Ashley Madison, the one known case of a breach exposing fake deletion
  • FTC, 2016 settlement; snippet
  • users paid $19 for “Full Delete”; the 2015 breach contained their data anyway
• I found no study of deletion requests after a breach, in either direction
  • neither “do people ask for deletion after a breach”
  • nor “do breaches show that deletion was not done”

tools from neighboring work

• DeBlasio et al., Tripwire, IMC 2017; read
  • registered unique fake accounts at “more than 2,300 sites”; a later login to the matching email account proved the site was breached; “detected 19 site compromises”
• Farooqi et al., CanaryTrap, PoPETs 2020; read
  • gave each of 1,024 Facebook apps a unique email address, then watched which addresses got unexpected mail
• Understanding Data Collection, Brokerage, and Spam in the Lead Marketing Ecosystem, arXiv 2026; snippet
  • 105 made-up profiles entered into marketing forms; opt-outs reduced calls and mail but none stopped them
  • closest prior work to idea 3, but about marketing forms, not people search sites or removal services
• Thomas et al., Ethical issues in research using datasets of illicit origin, IMC 2017; read
  • what a paper must justify before touching leaked data

research ideas

1: does removal last, and does the California portal work? my top pick

• question: after a real person asks for deletion, when does their profile leave people search sites, and does it come back?
• why now
  • brokers had to start processing portal requests on 1 Aug 2026
  • observe profiles before each new request to establish their starting state
  • my recommendation: compare independently observed accessibility with portal-reported status
• what to build: a crawler that looks a consenting person up on 20–50 people search sites every week and records what is shown
  • search by name, city, and old addresses, since a removed profile can return under a new URL
  • a person confirms each match, since He et al. show most automatic matches are wrong
• compare three groups of volunteers: state portal, one paid service, no request
  • people outside California are a natural no-portal group
• outcomes: share of profiles gone at 90 days, share back at 6 and 12 months, and how often the portal or dashboard says “deleted” while the profile is still up
  • 90 days is a chosen observation point, not an asserted deletion deadline
• risks
  • Tsudik’s group names this as future work
  • people search sites block crawlers
  • many people search sites claim their data is public and outside the law, so the portal may not touch what people care about most; that would itself be a finding
  • needs ethics board approval and volunteers
• cheap first step: 5 volunteers, 10 sites, 4 weeks, to learn whether crawling and matching are reliable

2: do breaches expose data that should have been deleted? the human’s idea

• I read “find out if exercised when data breach occur” two ways; I am not sure which was meant
  • a: a breach shows whether erasure really happened
  • b: people ask for erasure after a breach
• a, without touching stolen data
  • US states publish the breach notice letters companies must file, e.g. California, Maine, Massachusetts
  • read each letter and the company’s own statements: how many victims were former customers, rejected applicants, or closed accounts, and how old was the data?
  • from my memory, unchecked: AT&T 2024, T-Mobile 2021, and Optus 2022 all said former customers were affected
  • compare with the retention period the company’s privacy policy promised
  • output: “x% of breached records belonged to people who had already left”, a number I did not find anywhere
  • an LLM can label thousands of letters; a person checks a sample
  • weakness: letters are vague, so many will say nothing about how old the data was
• a, stronger but slower: planted accounts
  • combine Tripwire and Rupp et al.: make unique accounts at many services, ask for deletion, keep the email addresses alive for years
  • any later mail, login attempt, or appearance in Have I Been Pwned for a deleted account proves the data survived
  • catches services that would lie in answer to an access request
  • weakness: breaches are rare, so results take years; mail after deletion is the faster signal
• a, with surveys: repeat Mayer et al. and ask people whether they had closed the account before the breach date
  • cheap, but memory is unreliable
• b: do deletion requests rise after a breach?
  • big companies must publish yearly counts of delete requests under CCPA; registered brokers report theirs to the state registry
  • compare a breached company’s counts the year before and after, against similar companies with no breach
  • weakness: yearly numbers and few breached companies with published counts; I expect a weak result
  • finer signal: search interest in “delete [company] account” by week

3: does opting out leak your data?

• question: when you opt out, or a removal service does it for you, does the contact data you hand over get used or passed on?
• why it matters: opting out requires giving brokers your email, phone, and sometimes ID, and some removal services are tied to brokers
• method: made-up people, each with a unique email address and phone number per broker or service, never used anywhere else
  • any mail or call to that address shows that this broker used or shared it
  • van Kempen et al. already sent made-up identities to 322 brokers but with only two identities, so they could not tell who leaked
• easy to build: email addresses on our own domain, a few hundred forms, then wait
  • the LLM form-filling agent of Sun et al. shows the forms can be driven automatically
• risks
  • testing paid removal services needs real-looking customers and money
  • a null result is likely for big brokers; the interesting cases are the long tail of small sites
  • sending made-up requests costs brokers time; ethics review needed

smaller ideas

• map who owns which people search site, using shared hosting, page templates, and opt-out form backends
  • Take et al. found 4 groups behind 14 sites by hand; He et al. list 2,024 brokers
  • output: the shortest list of opt-outs that covers the most sites
• a free open opt-out agent, measured against paid services
  • He et al. ask for it; unpaid open-source attempts already exist, so the research part would be the measurement, which is idea 1
• where does a profile come back from?
  • plant a new fact about a consenting person in one source, e.g. a change of address form, and time its arrival at each broker
  • shows which upstream source to cut; slow and ethically touchy

what I would skip

• another one-shot count of how many brokers answer requests: done three times since 2024
• another short test of paid removal services: done twice
• anything that needs downloading breach dumps: legal and ethical cost is high and idea 2 does not need it

open checks before starting

• read the full Consumer Beware paper and the EDPB report; I only saw summaries
• search again for a DROP evaluation; one may appear any week
• confirm which states publish breach letters in bulk and how far back
• ask whether the human meant reading a or b of their breach idea

8 October measurement constraint

• CalPrivacy’s status page: “status updates can take up to 90 days to appear in DROP depending on data broker workflows”
  • quote checked through the primary page’s search-visible text
  • direct agency landing-page retrieval returned HTTP 403
• compare request time, independently observed accessibility, and provider-reported status separately
  • a pending status alone does not establish that a broker still exposes a record
  • broker-reported deletion alone does not establish downstream deletion or lasting removal
  • reappearance may be a new collection rather than retained old data
• the agency’s published reporting window is a measurement constraint
  • this update does not audit legal compliance

Last edited: 2026-10-08