Clarista White Paper · Building the Thinking System, Part Three
What it takes, what it costs and what it makes possible. Five beliefs keep wealth firms waiting on their own data. Each made sense ten years ago, and each is out of date.
Over dinner, a client of eleven years sets down her fork and says what has been on her mind. Her son graduated in May and is still looking for work. The entry-level jobs he trained for are the ones employers now hand to AI. Her mother turned eighty-eight in March, and the doctor has started to use the words "full-time care." She and her husband are paying for both. She wants to know one thing: can we carry this for five years without changing our own plans?
The advisor knows it is the right question. He also knows what answering it involves. The care costs need a place in the plan. The son's support comes out of a joint account at one custodian, while the cash that should fund it sits at another. The trust that pays her mother's bills is a PDF in the document vault. Last month's call about selling the rental property is in the notetaker, and her text confirming the sale is on his phone.
Each fact exists somewhere in the firm, each in a different system. The answer she gets on Monday depends on how much of it someone has time to piece together by hand.
Her situation is common. Pew Research Center found that 54 percent of Americans in their 40s have a parent aged 65 or older and are also raising a child or supporting a grown one. A private room in a nursing home cost a national median of $129,575 a year in 2025.
The rest of the picture. Care at home cost $80,080 a year in 2025. A woman who reaches 85 can expect to live seven more years on average. Recent college graduates face 5.6 percent unemployment against 4.2 percent for all workers, and Stanford researchers found a 13 percent relative drop in employment for early-career workers in the jobs most exposed to AI.
The question she asked is ordinary. The work behind a good answer is what this paper is about. It is written for the people who decide what a firm does about its data: a COO, a CTO, a head of operations, or a founder who still does all three jobs.
We hear five things in almost every first meeting. The first is the most common: in the 2026 T3/Inside Information software survey, 42.9 percent of advisors said they use an AI notetaker.
| What we hear | What is true today | Where this paper answers it |
|---|---|---|
| "We already use an AI notetaker. What else can AI really do for us?" | The notetaker is one source among nine. AI earns its keep when notes are read together with custody, planning, the CRM, calls and texts. | Putting It to Use |
| "Bringing our data together means hiring engineers and one to two years of spending. Is it worth it?" | Infrastructure is the small cost. The work is the cost, and most of it can now be automated: a source is reliable in about a month, and a first app can be in use in six to eight weeks. | Where the Time Goes and What It Costs |
| "A data lake, a data warehouse and a lakehouse are the same thing." | A lakehouse is a data lake with reliable tables, and the two are one choice. A warehouse differs from both in cost, in what it can hold and in how tied you are to one vendor. | Data Lake vs Data Warehouse |
| "We will need a connector for every vendor." | Connectors were needed when warehouses were wired to each vendor's format. Today data lands as it arrives and is shaped when it is read. | What It Takes to Make Data Ready for AI |
| "All our data has to be in one place before we can use any of it." | Data can be read where it sits, under one set of definitions and access rules. | What It Takes to Make Data Ready for AI |
Before deciding where data should live, list what you have. The RIAs and family offices we work with draw on the same nine kinds of source. The names of the vendors differ; the shapes of the data do not.
| Source | What it holds | How it arrives | How it names the client |
|---|---|---|---|
| Custody Schwab, Fidelity, Pershing | Positions, transactions, balances, cost basis | Daily files after the close, and APIs | Account number |
| Portfolio management Orion, Black Diamond, Addepar, Tamarac | Performance, models, billing, household groupings | Exports and APIs, daily | Portfolio or household ID |
| CRM Salesforce, Wealthbox, Redtail | Households, people, relationships, tasks | API, as people work | Contact and household ID |
| Meeting notetaker Jump, Zocks, Zoom | Transcripts, summaries, action items | After each meeting | Attendee name or email |
| Phone calls and text messages | Recordings, transcripts, messages | Archive exports, continuous | Phone number |
| Financial planning eMoney, RightCapital, MoneyGuide | Goals, cash needs, scenarios | Exports, when a plan changes | Plan or client ID |
| Alternatives | Commitments, capital calls, distributions, valuations | Notices and statements, often quarterly and late | Investor name and fund |
| Documents | Trust deeds, statements, agreements, tax returns | PDFs, when received | File name and folder |
| Market and reference data | Prices, security details | Daily feeds | Ticker or CUSIP |
Read the last column. Each system names the same family in its own way: an account number, a household, the people on a call, a phone number, a name on page one of a trust deed. Joining them is the first real piece of work.
Some sources arrive every night in tidy rows. Others arrive as conversations and documents, which hold facts no other system has, such as the rental sale mentioned on a call. Older tools could only hold rows, so many projects left these out. AI can now read a transcript or a PDF and pull out each fact with its source and date.
In wealth management, "data aggregation" often means account aggregation: pulling a client's held-away accounts, such as a 401(k) at an employer's provider or a bank account, through services like Plaid, Envestnet Yodlee or Morningstar ByAllAccounts, usually with the client's authorization. That is useful, and it is one input.
Firm data aggregation is the larger job. It brings together everything the firm itself holds about a family, including the conversations and documents, so it can be read as one picture. Account aggregation feeds it. This paper is about the larger job.
Three kinds of system can hold a firm's data. The terms are used loosely, and vendors use them to mean whatever they sell, so here they are in plain words.
A relational database, such as SQL Server or PostgreSQL, keeps data in tables with a fixed design, on one server where storage and computing power sit together. It is very good at reading and writing single records, which is why your CRM runs on one. You pay for the server whether anyone is using it or not.
A cloud data warehouse, such as Snowflake, BigQuery or Redshift, is built for analysis across large tables. Snowflake is a warehouse built on low-cost object storage, with storage and compute kept separate: your data sits in Snowflake's own compressed format, and you pay for computing power by the second while queries run. It is fast and well managed. Your data lives in the vendor's format, and the bill grows with how often the warehouse runs.
A data lake, such as Amazon S3 or Azure Data Lake Storage, is cloud storage for files of any kind: nightly custodian files, PDFs, transcripts, tables. It costs $0.023 a gigabyte a month on Amazon S3. A lakehouse is the same data lake with its tables kept in an open format, such as Apache Iceberg or Delta Lake, which adds safe updates and a history of every change. For a wealth firm the two are one choice: storage you own, in formats any engine can read. This paper treats them as one option.
| Option | Good at | You pay for | Who can read the data |
|---|---|---|---|
| Relational database | Apps that read and write single records | The server, all year | The database |
| Cloud data warehouse | Analysis across large tables | Compute while it runs, storage per terabyte | Mainly the warehouse |
| Data lake or lakehouse | Holding everything, including documents and conversations, at low cost, with reliable tables | Storage per gigabyte, queries per terabyte read | Any engine you choose |
A data lake answers questions through query engines that read its files directly, such as Trino, Amazon Athena, Azure Synapse serverless and BigQuery. You pay for the data each query reads: Athena charges $5 for each terabyte scanned. Much of a wealth firm's analysis, such as allocation by household, cash across custodians or fee calculations, can run this way, straight from low-cost storage, without loading a warehouse first.
Storage has become so cheap that you can keep everything, including every document and transcript, in storage you own, and choose the engine for each job. Choosing a warehouse used to mean choosing where your data lives for the next decade. It no longer has to.
Go back to the dinner. Answering her question draws on the eight capabilities this paper keeps returning to. Connecting the sources is the quick one. Four others carry most of the weight for her family.
Mastering and linking records. Her joint account at one custodian, her IRA at the other, the trust where she is trustee and her mother is the beneficiary, the CRM household and the plan that lists her son as a dependent all describe one family. Each uses a different identifier. The trust is titled in a way no system recognizes as hers. The call and the text belong to the household only once someone matches a phone number and an email to her record. Joining these is most of the work.
Defining terms and cataloging. "Liquid assets" has to mean one thing. Does it include the trust? The IRA, before or after tax? Cash promised to a capital call next month? Each firm answers differently, and the answer has to be written down once, so every report and every AI answer uses it.
Quality monitoring and reconciliation. Positions are checked against the custodian's own file. Cost basis is missing for lots transferred in from a previous firm. The private fund's last valuation is a quarter old. A feed arrives late, or a vendor changes a file layout. Each gap is flagged with its source and date, and caught before anyone builds on it.
Extracting facts from documents and transcripts. The rental sale becomes a dated event, taken from the call transcript and confirmed by her text. The trust's distribution terms come out of the PDF with a page reference. Each fact keeps its source, so the advisor can open it and check.
The other three decide who may see her family's data, how each figure can be traced, and what it costs to run: access controls and data masking, lineage and audit records, and cost and performance tuning.
AI now does much of this work, which is why it takes weeks where it used to take years. It pulls facts from documents, proposes which records belong to one family, drafts definitions from your existing reports and notices values that move more than they should. Your people confirm each match and each definition, because the firm's language belongs to the firm.
For one source, connecting and landing the data takes a couple of days once the vendor grants access, or one to two weeks with the vendor's paperwork. The month after that goes into making it reliable. Several sources run side by side, and a first app is typically in use in six to eight weeks.
Look at Figure 4 again when a vendor tells you how fast their connectors are. Connection is the part that takes days. The month is the work that decides whether anyone can rely on the result, and it has to be done by someone.
Cost is the question we are asked most, and the one least often answered with numbers. Here is a reference firm, the data it holds, and what three common set-ups would cost to run it at published list prices.
| Assumption | Value |
|---|---|
| Assets under management | $3 billion |
| Households | 1,500 |
| Accounts | 6,000 (4 for each household) |
| History kept | 10 years |
| Positions | 40 for each account, every trading day |
| Transactions | 150 for each account a year |
| Documents | 300 for each household, 0.5 MB each |
| Meeting transcripts | 4 meetings a household a year |
| Data held | Size |
|---|---|
| Positions, ten years of daily history | 90.7 GB |
| Transactions, ten years | 1.8 GB |
| Documents | 225.0 GB |
| Meeting transcripts | 3.0 GB |
| Calls and text messages | 5.0 GB |
| CRM, planning, alternatives and reference data | 5.0 GB |
| Total | 331 GB |
All of that comes to roughly 331 gigabytes. Two thirds of it is documents. Ten years of daily positions for 6,000 accounts, which sounds large, is 605 million rows and fits in under 100 gigabytes.
Each set-up stores the same data and serves the same questions. Documents sit in object storage in every case, because a database is a poor place to keep PDFs.
Assumptions: 0.25 TB of tables and 0.75 TB of documents, rounded up to leave room for growth. Amazon S3 $0.023 per GB a month; Athena $5 per TB scanned, 10 TB a month; RDS PostgreSQL db.m6i.xlarge $0.356 an hour in two zones, with gp3 storage at $0.115 per GB a month; Snowflake Enterprise $3.00 a credit, X-Small warehouse for 6 hours a day, storage $23 per TB a month.
Storage and compute are the small line. The larger one is the work that makes the data something an advisor can rely on, show a client and put to use: the eight capabilities in Figure 4, from connecting sources and linking records into households to masking tax IDs, keeping an audit trail and holding compute bills down. Whichever set-up you choose, someone has to configure and run each of them.
Every capability costs $32,584 to $73,314 a year in engineering time, and the smallest of them costs more than the most expensive infrastructure. The split across capabilities is our estimate from client work; your own mix will depend on your sources.
The total assumes two data engineers, the smallest in-house team that can cover holidays and turnover. The median US database architect earns $139,500 a year, and benefits bring the cost to the employer to $203,650. Two of them cost $407,299 a year, before recruiting, management and the months it takes to find them.
That is the real decision behind the second belief. The question is less "can we afford a data platform?" and more "who will do the work, and how much of it can be automated?" Platforms such as Clarista, and packaged wealth data platforms, replace most of the people line with a subscription and a set-up fee. Ask any vendor, including us, to put their price next to the $407,299.
About these numbers. They are list prices, US East region, from the vendors' published rates as of October 2026, with no reserved-capacity or negotiated discounts. They leave out data transfer, backups and the cost of building reports. Pay figures are national medians; pay in New York or Boston is higher. Use them as a starting point, then ask for a quote built on your own volumes.
A data lake holds the data. Point an AI assistant at it and the assistant still cannot tell which accounts belong to the family from dinner, what "liquid assets" means at your firm, or which advisor may see the trust. Six building blocks sit between the storage and the people and tools that use it. Figure 7 shows how they fit together.
| Building block | What it does | In a wealth firm |
|---|---|---|
| One query layer | Reads each store where it sits: the lake, an existing warehouse, the CRM's database. New feeds land as they arrive and are shaped when they are read. | A custodian adds a column to its daily file. The feed keeps landing, and only the definition that uses the column changes. |
| Semantics and catalog | Defines each term once in the firm's words, and describes every table and field. | "Liquid assets" means the same thing in the quarterly report, the meeting pack and an answer from Claude. |
| Firm context | Links records into households and entities, and holds the figures the firm works out from them. | The trust where she is trustee is tied to her household, and household liquidity is calculated by the firm's rule, not guessed by a model. |
| Access controls | Sets rules once, by role, household or field, and applies them to every store and every way in, under the identity of the person asking. | An associate advisor sees only the households she serves. Tax IDs are masked outside operations, in Excel and in ChatGPT alike. |
| Audit log | Records who asked what, through which tool, when, and what came back. | An examiner asks what an advisor saw about a client on a given date. The answer is a record, not a recollection. |
| Ways in | MCP for AI assistants, REST APIs for apps, JDBC and ODBC for BI tools and Excel, Python and SQL for analysts. | The same household figure appears in a Power BI dashboard, a custom app and a Claude answer. |
Older warehouses needed every feed reshaped to fit fixed tables before it could load, so each vendor needed its own connector, and a renamed column stopped the load until someone rewrote it. Today feeds land as they arrive and are shaped when they are read, an approach called schema on read. Data can also be read where it already sits, through one query layer across several stores, called data federation. Copying becomes a choice made source by source, for history or speed.
Your data does not need to sit in one place. It needs to be linked, defined and checked, wherever it sits.
Our paper What Is an AI Agent Harness? put it simply: an agent is a model plus a harness, and the harness decides what the model sees, remembers, can do and may not do. The building blocks in Figure 7 are the data half of that harness. The firm's linked data is the agent's memory, semantics and firm context decide what it sees, and access controls and the audit log are its guardrails. Models will keep changing; these blocks stay with the firm.
MCP deserves one note. It is a standard way for AI assistants to connect to data, and it says nothing about which rows or fields a person may see. That has to be enforced in the layer underneath, as The Door AI Walks Through explains.
The SEC's amendments to Regulation S-P, adopted in May 2024, now apply to every registered adviser. Firms with $1.5 billion or more in assets had to comply by December 3, 2025, and smaller firms by June 3, 2026. Each firm needs a written program to detect, respond to and recover from unauthorized access to customer information, and must tell affected clients within 30 days. The rule also reaches the firm's vendors. Advisers must oversee every service provider with access to customer information, through due diligence and ongoing monitoring, make sure each one reports a breach within 72 hours, and keep records of that oversight. Advisers coming out of recent SEC examinations tell us the questions now center on vendor risk: technology, business continuity, recovery and the use of AI. This is why access controls and the audit log are building blocks rather than extras. Once a data platform and AI tools read client data, the firm has to show who could see what, who did see it, and how quickly it would hear of a problem.
A data lake on its own answers nothing. Here is what trusted data makes possible, starting with the question from dinner. Each outcome names the question in an advisor's words, the sources it joins and what the firm gets.
Sources joined: both custodians, the trust document, the plan, the CRM, the call transcript and the text. The firm gets a five-year view of the family's cash: care paid from the trust, support for the son, the rental sale, taxes and the plan's own goals, each figure opening to its record, and a place to record the reasoning behind the recommendation.
"Which households should we review for tax moves before year end?" Sources joined: cost basis and tax lots from custody, the portfolio system, the plan, meeting notes, and K-1s and tax returns read from documents. The firm gets a list of households worth a closer look, each with its reason: losses that could offset gains, a lower-income year that may suit a Roth conversion, appreciated positions that could fund a planned gift, or a K-1 still outstanding. With every account linked, a sale in the taxable account is checked against purchases in the IRA for wash sales. Each candidate goes to the advisor and the client's tax professional to decide.
"What do I need to know before Thursday's review?" Sources joined: the portfolio, custody, the plan, the CRM, the last meeting's notes, calls and texts since then, and market moves in the family's holdings. The firm gets a one-page pack: what changed, what was promised and not yet done, and the questions worth asking. The notetaker is one input; the pack is what the advisor actually needed.
"Which families have capital calls due this quarter, and is the cash where it needs to be?" Sources joined: capital call notices read from documents, commitment records, custody balances and the plan. The firm sees each call next to the cash that will fund it, weeks before the due date.
"Did we bill every account correctly this quarter?" Sources joined: custody values, the portfolio system's billing groups, fee schedules from the CRM and the signed agreements. Every exception comes with the figures that produced it.
"Did a client ask for something by text or on a call that we have not acted on?" Sources joined: message and call transcripts, the CRM and custody activity. Requests are matched to what happened next, and open ones are listed for review.
A multi-billion dollar advisory firm we work with kept roughly 700 spreadsheets, one or more for each household, holding managed and oversight accounts, cash needs, allocation at household and account level, and stated risk against actual risk. Linking those pieces into one view of each household removed the reconciliation. That view then became the input to trading, which moved from an entirely manual process to one-click generation followed by review.
The same work has paid off beyond wealth management. Edelweiss, an alternatives manager running hundreds of funds across several custodians and administrators, cut its data processing time by 95 percent. Reconciliations that took seven to ten days now take hours, and over 70 percent of flows reconcile with no one touching them. Connecting its sources took four months, where a traditional build would take 12 to 18. Leerink Partners, a healthcare investment bank, had a cloud data lake running in four months, with 22 two-way interfaces keeping its legacy systems and new CRM in step every two minutes. The full stories are in our case studies.
This is why data comes first in the series. Part One, The Thinking System, described a firm whose knowledge lives in a system instead of in a few people's heads. Part Two, Building the Thinking System, set out the layers that make it work: capturing each client event, recording the judgment behind each recommendation, and agents that reason from both. Each layer stands on this one. An event can be captured only when it is tied to the right household, a recommendation recorded only against figures the firm has defined, and an agent can reason only from data it can trace to its source. Productivity agents, like the meeting pack, are useful from the first month. Reasoning agents grow more useful as the firm's record of judgment builds.
On Monday the advisor calls her back with the five-year view. It shows care paid from the trust, her son's support, the rental sale and the tax on it, and where the plan has room and where it does not. Every figure opens to its source: the custodian's file, the page of the trust, the call where she mentioned the sale. He records why he recommends what he does, so the next conversation starts from there. That is wealth data aggregation doing its job.
Which three questions do our principals ask that take more than a day to answer?
Which systems hold the facts behind them, and who looks after each one?
Where do we keep the same fact in two places, and which one do we trust?
Which conversations and documents hold facts that no system has?
What would we stop doing by hand if one of those questions took a minute?
Where does our data sit, and can we query it ourselves, without you?
Who links accounts into households and defines our terms, and is that work included?
How is each number checked against its source? Show us the source and date on one figure.
Can our AI tools read the data under each person's own permissions? Show us the log of one question.
Can we build and own apps on the same platform, and how soon can the first one be in use?
How long is the contract, and what do we keep if we leave?
Do you hold SOC 2 Type II and ISO 27001, and will you share the reports each year?
Who on your side can reach our client data, and how is that access controlled and logged?
Is breach notice to us within 72 hours written into the contract?
What are your service levels for business-critical work, such as feeds before trading and billing at quarter end, and what happens when one is missed?
How do you recover from an outage, and how long does it take?
Which AI models read our data, where do they run, and is our data used to train them?
Clarista is one framework that works with the storage you already have or choose: a data lake, a lakehouse, a warehouse such as Snowflake, or the databases you already run. Your data stays in your own cloud account, and Clarista configures all eight capabilities from Figure 6 on top of it. AI does the routine work; your people decide what only the firm can decide.
| Capability | What Clarista does | What your firm decides |
|---|---|---|
| Connecting sources and handling vendor changes | Sourcing workflows land custody, portfolio, CRM, planning, notetaker and document feeds in your lake. A change in a vendor's file is caught before it reaches a report. | Which sources come first |
| Mastering and linking records | AI proposes which accounts, people, trusts and entities belong to each household, with the evidence for each match. | Confirms each match |
| Defining terms and cataloging | AI drafts definitions from the way your reports already calculate figures, and describes every table and field. Each definition is held once and used everywhere. | Approves the firm's language |
| Extracting facts from documents and transcripts | Trust deeds, capital call notices, K-1s and meeting transcripts become dated facts, each with its page or timestamp. | Confirms facts before they drive a recommendation |
| Quality monitoring and reconciliation | Checks and alerts on every feed: positions against the custodian, stale valuations, missing cost basis and values that move more than they should. | The tolerances for each check |
| Access controls and data masking | Rules are set once in the query layer and applied under each person's identity, to every store and every way in: MCP, REST APIs, JDBC and ODBC. | Who sees which households and fields |
| Lineage and audit records | Every query is logged with who asked, through which tool and when. Every figure opens to the records it came from. | How long logs are kept, and where |
| Cost and performance tuning | A query cluster sized for your users, reading from low-cost storage instead of an always-on warehouse. | The budget and the response times |
Each source is connected in days and reliable in about a month, with several sources running side by side. The first app is built alongside and is typically in use in six to eight weeks.
You own the data and the apps. The lake sits in your cloud account, and you keep full SQL access to it, with or without Clarista.
Beyond the eight capabilities. The same platform delivers the AI-ready building blocks in Figure 7: one query layer, the semantic layer and firm context, and access through MCP, REST APIs, JDBC and ODBC under one set of access rules and one audit log. Apps and agents built on the data run on the Clarista Hosting Platform, with single sign-on, managed secrets, automated deployment and security monitoring. This paper covers both only briefly; What Is an AI Agent Harness? describes them in full.
Clarista holds SOC 2 Type II attestation and ISO 27001 certification. See the data lakehouse for RIAs and family offices for more detail, or book a demo and bring the question from your own dinner table.
Wealth data aggregation brings together the data a wealth firm holds about each family, from custodians, portfolio management, CRM, planning, alternatives, meeting notes, calls, texts and documents, so it can be read as one linked picture. Pulling the data in is the quick part. The work is linking records into households, defining terms, checking values against their sources and watching each feed.
Account aggregation pulls a client's held-away accounts, such as a 401(k) or a bank account, through services like Plaid, Envestnet Yodlee or Morningstar ByAllAccounts. Firm data aggregation brings together everything the firm itself holds about a family, including conversations and documents. Account aggregation is one input to it.
Most RIAs are best served by keeping their data in a data lake or lakehouse they own, because storage is cheap and it can hold documents and transcripts as well as tables, and then choosing the query engine for each job. A warehouse is a good engine for heavy analysis, and it can read from or sit beside the lake.
Only if your analysis needs a warehouse engine. Snowflake is a warehouse built on low-cost object storage, with storage and compute kept separate. Many wealth firm questions can be answered by query engines that read files directly in a data lake, at lower cost. A firm that already owns Snowflake can keep it and read it alongside other stores.
A lakehouse is a data lake whose tables are kept in an open format, such as Apache Iceberg or Delta Lake, which adds safe updates and a history of every change. Both keep data in low-cost storage you own, in formats that any engine can read. For most wealth firms they are one choice, compared with a data warehouse.
Storage alone is not enough. AI-ready data needs one query layer that reads each store where it sits, terms defined once in the firm's words, records linked into households, access controls set once and applied to every tool under each person's identity, an audit log of every question, and ways in for AI assistants (MCP), apps (REST APIs), BI tools (JDBC and ODBC) and analysts (Python and SQL).
Schema on read means data is stored as it arrives and its structure is applied when it is read, through definitions that can change without reloading. It is why a new vendor feed can be landed in days instead of waiting for a custom connector.
Data federation is a single query layer that reads across several stores at once, such as a data lake, an existing warehouse and a CRM database, under one set of definitions and access rules. It means data does not have to be copied into one place before it can be used.
For a $3 billion RIA with 1,500 households and ten years of history, infrastructure costs $876 to $7,134 a year at list prices, depending on the set-up. The larger cost is people: two data engineers cost $407,299 a year including benefits.
Connecting a source and landing its data takes a couple of days once the vendor grants access, or one to two weeks with the vendor's paperwork. Making it reliable, by learning what it holds, finding gaps, linking records and setting quality checks, takes about a month. A first app can be in use in six to eight weeks.
Choose which categories of cookies and tracking technologies you allow. Strictly necessary cookies cannot be disabled. Read our full Cookie Policy.
Required for security, authentication, fraud prevention, load balancing, and storing your consent preferences. Always on.
Remember choices you make to provide a personalized experience (language, region, theme, saved filters).
Help us understand how the site is used (Google Analytics, Vercel Analytics, Microsoft Clarity, etc.).
Used for conversion tracking, retargeting, and personalized ads (LinkedIn, Meta, Google Ads, Bing, etc.).