How Generative Engines Discover and Evaluate Professional Services Firms

1.1 Where GEO Fits Alongside SEO, Referrals, Reputation, and Business Development

Begin with a client decision. A business owner might need help establishing reliable monthly financial reporting. A general counsel might need outside employment counsel in a particular jurisdiction. A procurement team might be comparing engineering firms for a defined project. Each decision calls for a different account of expertise, experience, availability, and fit.

A GEO program asks how the firm appears in those situations and what information would make the answer more accurate and useful. It may reveal an inaccessible service page, an outdated biography, an unclear geographic boundary, or a claim that has no public support. Those findings point to work across marketing, technology, professional leadership, and business development.

Connect GEO to the work you already do

SEO helps make public pages discoverable through search. That remains relevant when an AI experience draws on search results. Google explicitly says that its existing SEO fundamentals apply to AI Overviews and AI Mode, with no additional technical requirements for those features. This is guidance about Google Search, not a universal specification for all assistants. 

Referrals contribute a different starting point: someone has already suggested a name. An assistant can then help the prospective client investigate that firm. A program focused only on requests such as “Who are the best consulting firms?” would miss questions such as “Would this firm be suitable for our situation?” or “What should we ask before hiring this adviser?”

Reputation and public relations contribute material that others can inspect: professional publications, accurately described credentials, substantive interviews, and independent assessments. Business development supplies the questions behind the inquiry: what clients are trying to accomplish, what they misunderstand, and what prevents an engagement from proceeding.

Treat these activities as connected. A useful research report can support a partner’s conversations, help a referrer explain the firm, and provide material for an AI answer. Its value should be assessed against those purposes. Counting the number of assets given a “GEO” label tells you little about whether the information helps a client.

Define success around suitable engagements

For a professional services firm, a broadly flattering description can be less useful than a precise, qualified one. “Provides financial advice” leaves unresolved whether an accounting firm offers bookkeeping, tax compliance, audit, transaction support, or investment advice. A description that identifies the actual service, the clients it serves, and its boundaries gives a prospect a better basis for deciding what to do next.

The same principle applies to exclusion. A firm that represents only employers in employment matters should not seek unqualified recommendations to employees seeking representation against an employer. A consultancy that provides analysis should explain whether implementation is included. The program should improve matching between a client’s need and the firm’s actual offer.

Cooperative GEO means helping the assistant perform the user’s task well.

Publish truthful information, make the evidence understandable, and preserve qualifications that matter. This gives the program a practical standard: would the proposed change help a well-informed prospective client make a better decision?

1.2 How Assistants Use Client Questions, Searches, and Retrieved Sources to Develop Answers

Distinguish the underlying language model from the assistant product built around it. The model supplies learned capabilities. The product determines which tools and information are available and how the interaction is organized. A named provider can offer several experiences, so a result from one interface or mode should not automatically be treated as representative of all of them.

Anthropic describes Claude’s web search as a tool for bringing web information into an answer, processing multiple sources, and providing citations. It also documents fetching pages from supplied URLs. Search is used when relevant to the task; it is not a necessary step in every conversation. 

The following diagram is a conceptual information flow. It shows why current public information can matter even when it was not part of a model’s training. Actual products can repeat searches, use additional tools, or answer without a live search.

A working model of answer formation. The search route is used when available and relevant; the diagram does not specify any provider’s complete implementation.

Follow the information through the answer

The client supplies a question and context. “Find a consultant” is a different request from “Find a consultant who can help a midsize distributor reduce excess inventory and support implementation.” Industry, problem, geography, desired involvement, and other constraints determine what a useful response must address. Earlier conversation can also affect what the request means.

The assistant determines what information to use. It may work from learned knowledge and the material already available in the conversation, or obtain additional information through its tools. Where the experience supports them, a client might supply a firm name, a website link, or a proposal. Those inputs change what is available for that particular assessment.

Search can expand the original request. OpenAI documents that ChatGPT search may rewrite a question into targeted searches and issue further queries after examining results. Google describes a similar possibility, called query fan-out, in its AI Search features. 

For the distributor’s request, plausible searches could concern inventory improvement, distribution-sector experience, and implementation support. These are illustrative possibilities, not a record of what a particular assistant searched. The practical implication is to explain the substantive dimensions of a service clearly, so relevant information exists for the different parts of a client’s question.

Source material becomes evidence for the response. A service page may describe the offer; a biography may identify relevant experience; an approved case study may demonstrate the firm’s role in a project. An outside source may corroborate a particular fact. Each source has a scope. A professional directory entry that verifies a credential does not, by itself, verify every commercial claim on the firm’s website.

The assistant generates an answer. It may summarize an issue, describe a firm, assemble a shortlist, make a comparison, or ask the client to clarify the need. The answer may contain citations, qualifications, and errors. A coherent explanation should still be checked against the evidence.

The client decides what to do next. They might investigate a source, ask another question, contact a firm, or decide that no engagement is necessary. Seeing the firm in an answer establishes only that appearance. What the client subsequently does is a separate business question.

Consulting example

A CFO receives a referral to a boutique consulting firm and asks whether it would be suitable for a working-capital project at a distributor. The homepage promises “transformational impact.” Elsewhere, a practice page describes inventory analysis and receivables management, a partner biography identifies distribution experience, and an approved case study explains the firm’s implementation role.

Those specific materials could support a useful answer: the firm appears relevant to the stated problem, with evidence of particular capabilities; the CFO should confirm team availability, scope, and how implementation would be staffed. The generic homepage claim supplies little basis for that assessment.

The GEO opportunity is to make the relevant account of the firm accessible, consistent, and supported. The client relationship began with a human referral. AI still has a role in helping the buyer evaluate it.

1.3 The Difference Between Model Training and Current Information Retrieval

The phrase “get our information into the LLM” can refer to different processes. Separate them before commissioning work. Publishing a page, having an assistant retrieve it, and changing a model through training are different events.

Training changes the model’s learned parameters

During training, a model’s numerical parameters are adjusted as it learns patterns from information. OpenAI describes sources that include public internet material, partner information, and material provided or generated by users, trainers, and researchers. This helps explain how a model can respond without conducting a fresh web search. 

For the firm, learned knowledge can create a difficult maintenance problem. You cannot establish from a confident answer alone which historical source shaped a statement, or assume that a revised webpage has changed the model’s parameters. An answer may reproduce an outdated description even when the firm’s current offer has changed.

Retrieval supplies information for the current task

Retrieval brings material into the assistant’s working context. A recently published service description can therefore inform an answer if the system can obtain it and chooses to use it. This does not require that the description first become part of the underlying model through training. Anthropic’s web-search documentation describes this use of current web content to inform responses. 

Retrieval creates a practical route for current facts, but the information still needs to be available, relevant, and correctly interpreted. Searching now does not make every source current. A search can find an old biography, a stale directory listing, or a page that accurately describes only one part of the practice.

Information supplied directly by the user is another input to consider. If you paste a detailed firm description into a conversation, the resulting answer is an assessment with that material available. It does not demonstrate how the assistant would describe the firm without it. Similarly, correcting an answer in your own chat is not evidence that other users’ answers have changed.

Accounting example

An accounting firm historically focused on tax preparation and has recently introduced outsourced monthly accounting. A new announcement mentions the service, but the main service page and several outside profiles still describe a tax-only practice.

An answer using the announcement might mention the new offer. One using older descriptions might omit it. Even the new announcement may leave questions unanswered: what does monthly accounting include, which clients are eligible, and who delivers the work?

The immediate task is to establish a clear, current public account of the service and resolve conflicting descriptions. That gives a retrieving system better evidence to use. It does not establish when a model’s learned knowledge will change or guarantee that the next answer will include the service.

Check access policies by purpose and provider

OpenAI documents independent controls for OAI-SearchBot, used for search, and GPTBot, used to collect material that may contribute to model training. A website can allow the former while disallowing the latter. This illustrates why search access and training permissions should be examined separately; the exact controls differ by provider. 

Chapter 10 covers implementation. At this stage, ask your website owner to identify the existing policy and any uncertainty. Keep approved public information accessible according to the firm’s decisions, and protect private client material with appropriate access controls. The program does not need confidential records to be made public simply to explain the firm’s expertise.

1.4 What It Means to Be Mentioned, Cited, Described Favorably, or Recommended

“We showed up in the answer” is the beginning of an observation. To understand its significance, classify what actually happened.

These outcomes can occur together or separately. A firm might be named in a warning. Its article might be cited without its name appearing in the prose. It might appear on a shortlist with a qualification about geographic coverage. Preserve those distinctions when recording the answer.

OpenAI warns that search results and citations can be incomplete, outdated, or incorrect. Open the cited page and check whether it supports the relevant statement. A citation provides something to inspect; its presence does not settle the question of accuracy. 

Nor should visible citations be treated as a complete account of answer formation. OpenAI’s API documentation, for example, distinguishes inline citations from a larger set of consulted URLs. The operating lesson is to record what you can observe and avoid inferring a full sourcing history from the links displayed. 

Law firm example

A law firm publishes an explanation of employment disputes in a particular jurisdiction. An assistant cites the article when answering an employee’s question about understanding a dispute and finding representation. The firm’s published service description states that this practice accepts only employer-side engagements.

The article can be relevant to the explanation even though the firm is unsuitable for the employee’s request for representation. A useful answer could draw on the published explanation while suggesting that the reader investigate employee-side counsel. Converting the citation into an unqualified recommendation would confuse subject knowledge with client fit.

For the firm’s GEO program, the appropriate outcome is an accurate description of its practice. A citation in an educational answer and a recommendation to a suitable employer are different results, both of which can be recorded without treating an unsuitable inquiry as a success.

Evaluate the substance of favorability

A positive adjective is a weak basis for judging an answer. “Leading,” “trusted,” and “excellent” tell the reader little unless the answer explains what supports the assessment. More useful language connects a capability with evidence and a relevant situation.

For example, “The firm describes experience supporting acquisitions in this industry, including the following engagements” gives a buyer a claim to investigate. “The best adviser for any transaction” obscures differences in scope, team, sector, and project requirements. Your review should examine whether the wording is justified and useful, not merely whether it sounds complimentary.

Likewise, a qualification such as “confirm the team’s capacity” need not be unfavorable. Public evidence of experience cannot establish whether the required people are available for a particular start date. Preserving that uncertainty can make the answer more useful.

A recommendation also needs context. Appearing in a list of possible providers is different from being identified as the preferred choice after a detailed comparison. The order of a list is not automatically a search ranking. Record the surrounding question and wording before attaching significance to a position.

1.5 Distinguishing the Visibility of a Firm, Practice, Office, and Individual Professional

Professional services information concerns several related subjects. A firm’s name may cover multiple practices and offices. An individual may have a career that spans several employers. A network may include independently owned member firms. Your public information should make those relationships explicit.

The following four levels are a useful way to organize the work. They are this playbook’s information model, not a claim that every AI provider uses the same database structure or ranking formula.

A general firm description should not imply that every office delivers every service. A biography should not imply that all of a professional’s past engagements were performed by the current firm. A network affiliation should identify the relationship rather than leave the reader to assume common ownership or a single delivery team.

Engineering consulting example

A senior engineer joins a young consultancy after working on a major transport project at a previous employer. Her biography accurately identifies the project, her responsibilities, the employer at the time, and her current position.

An answer saying that the new consultancy delivered the transport project has assigned the experience to the wrong organization. The engineer’s experience remains relevant. It does not establish the new firm’s project history, the contribution of its current team, or the scope of work it can presently undertake.

This is also a professional communication issue. The NSPE Code of Ethics addresses accurate qualifications and warns against exaggerating responsibility for prior assignments or misrepresenting past accomplishments. That is a professional society’s code; applicable requirements should be checked in the relevant context. 

The same information problem arises when consultants, accountants, and lawyers change firms. The useful account preserves who performed the work, in what role, for which organization, and when. Removing those relationships can make an otherwise factual project list misleading.

Attach each important claim to the right subject

Before publishing a claim, ask who it describes. Is “twenty years of experience” the age of the firm, the career of one partner, or a combined figure for several people? Does “national coverage” mean staffed offices, remote service delivery, or access to independent affiliates? Does a credential belong to the firm, one office, or an individual?

Where the distinction affects suitability, put it in the explanation. Clear biographies, practice descriptions, professional records, and project accounts give readers material to check. Their usefulness does not depend on claiming that each is a known ranking factor for every AI system.

For later monitoring, keep these levels distinguishable. A correct description of the firm’s overall services can coexist with an outdated account of a departed partner. A practice can be visible for one kind of engagement and absent for another. A single firmwide score would conceal those differences.

1.6 Why Answers Vary, What Firms Can Influence, and Where Control Ends

An AI answer is an observation under particular conditions. Changes in the question, context, tools, available information, or generation process can produce a different result. A GEO program needs to distinguish those changes from a real improvement in how the firm is represented.

OpenAI documents that location and relevant saved memories may influence search queries. It also explains that model generation contains an element of randomness. Google states that AI Overviews and AI Mode can use different models and techniques, producing different responses and links. 

Consider three requests: “Explain the options for improving working capital,” “Find advisers for a working-capital project,” and “Assess this firm’s suitability for our working-capital project.” The first may be answered usefully without naming providers. The second asks for discovery. The third asks for an assessment of a known candidate. Different appearances across these requests can reflect different tasks.

Even apparently small qualifications can change the appropriate answer. Requiring implementation support narrows a consulting inquiry. Specifying employer-side representation changes a legal inquiry. Asking for an audit changes an accounting inquiry. Your program should preserve those distinctions rather than seeking the same favorable result in every test.

Observe before concluding

Save the question and answer, the date, the experience used, whether search was involved where observable, and relevant context. Repeat important observations before declaring a pattern. Detailed sampling and metrics belong in Chapters 5 and 6; the principle begins here.

If a firm appears in ten of fifty test answers, its mention rate in that sample is 20 percent. That calculation does not establish that it appears in 20 percent of all relevant client conversations. The sample’s questions and conditions determine what the result describes.

A verified, successful crawler fetch shows that the crawler could obtain the page; it does not establish that a particular answer used it. A retrieved page is not necessarily cited. A citation is not necessarily a recommendation. OpenAI explicitly states that placement in ChatGPT search is not guaranteed. 

Separate control, influence, and observation

• Direct control: the firm’s own descriptions, public pages, factual claims, permission records and publication approvals, publishing process, and technical configuration.
• Influence: outside profiles, editorial coverage, professional participation, and correction requests. Independent publishers decide what they publish.
• Outcomes to observe: which sources an assistant selects, how it describes the firm, and what prospective clients do afterward. These require evidence and cannot simply be specified in a content brief.

This distinction helps you commission concrete work. A supplier can commit to verifying facts, improving a page, fixing an access problem, and documenting observations. A promise of a particular independent recommendation requires a different level of scrutiny.

Early GEO research found that changes to source content could affect its prominence in generated answers under tested conditions. In the original GEO paper’s Perplexity evaluation, researchers supplied source text as files and constrained answers to those files. Those findings do not establish the result of editing a live firm website and waiting for organic discovery, recommendations, or client engagements. 

Treat such research as a source of hypotheses. Ask what changed, which experience was tested, what outcome was measured, and what comparison supports the conclusion. A method that improves the presentation of supplied information has not necessarily solved the separate problem of getting that information discovered.

For your own program, start with improvements that are defensible to a prospective client: clearer service boundaries, accurate expert profiles, accessible explanations, and evidence matched to the claims. Then test how the relevant AI experiences respond. The work should remain useful and truthful even when the measured effect is uncertain.

Practical Output: Build a One-Page Information Flow Map

Choose one service and one plausible client decision. You do not need a monitoring subscription or a complete content audit for this exercise. The aim is to inspect one information path and identify the questions your program will need to resolve.

1. State the decision. Identify who needs help, what engagement they are considering, and one or two conditions that affect suitability.
2. Write a realistic question. Use the language a client or referrer might use. A named-firm assessment is as valid as a request to discover providers.
3. Identify a small set of public sources. Start with a service page, an expert biography, any publishable evidence of relevant work, and a pertinent outside source where one exists. Record actual URLs.
4. Extract supported facts. Note what each source establishes, which entity it concerns, its date, and any qualification needed to interpret it.
5. Describe the answer those facts could support. Include a factual explanation of fit, important limitations, and what still needs direct confirmation.
6. Mark gaps and any observations. Identify missing, inconsistent, outdated, or ambiguously attributed information. If you inspect an actual AI answer, save it and distinguish its visible citations from sources you merely think would be useful.

Keep the completed map to one page. Use the following fields as a template:

In the accounting example, the map might show that the new service announcement is available but that the main practice description is stale and delivery responsibilities are unclear. Those are concrete information problems. You do not need to claim that they explain every AI answer before deciding that they deserve attention.