GEO for B2B services: agencies, consultancies, firms

The AI research habit was documented first for software, but the data now covers business buying broadly. Semrush surveyed more than 600 U.S. business professionals on how AI shapes B2B purchasing and found 66 percent use AI to explore possible solutions, 61 percent to compare vendors directly, 59 percent to understand a problem more deeply and 53 percent to ask outright for recommendations. Forrester's 2026 buyer journey research puts ChatGPT specifically inside vendor evaluation for 72 percent of B2B buyers. Every one of those behaviors applies with full force to hiring an agency, a consultancy or a firm, because choosing a services provider is precisely the ambiguous, comparison-heavy, trust-poor decision people bring to an assistant.

Against that usage stands the visibility gap: the same research wave found roughly half of brands earning zero AI citations. For services the gap is structural rather than accidental. Firms accumulate their proof in private, in decks, references and dinners, while engines can only cite what is public and crawlable. A twenty-year consultancy with a two-page website is, to a generative engine, a rumor.

The stakes carry over from the adjacent software data. G2's March 2026 survey found 69 percent of buyers chose a different vendor than they originally planned because of AI chatbot guidance, and a third bought from a brand they had never heard of before an assistant named it. Services purchases run on referral networks that took decades to build, and the assistant now functions as a referral source that every prospect consults privately, before the first call, with no one from your firm in the room. Being absent from that referral is survivable today and quietly compounding against you every quarter it persists.

The upside of the gap is that it is unclaimed ground. In most services niches nobody has done the GEO work yet, so the firm that externalizes its evidence first tends to become the recurring answer. The foundations of the discipline are covered in what is generative engine optimization; what follows is the services-specific application.

Services questions carry three qualifiers products rarely get: a niche, a stage and often a geography. Nobody asks for the best consultancy in the world; they ask for the boutique that has done this exact project for a company like theirs. That specificity is your friend, because specific questions have thin competition and are winnable by firms of any size.

When an engine compares software it has spec sheets, pricing pages and thousands of structured reviews. When it compares consultancies it has none of that, so it reaches for proxies: who appears on the directory pages and published rankings that already rank for "best X firms," whose case studies contain concrete numbers, whose people are quoted in credible places, and who the community threads recommend. 5W Research's citation data shows how much engines lean on third-party corroboration generally, with Wikipedia and Reddit together over a quarter of U.S. ChatGPT citations; in services categories, directories and rankings play the role review sites play for SaaS.

This is why the directory work matters despite feeling old-fashioned. Clutch-style platforms for agencies and dev shops, legal and financial directories for firms, industry association listings for consultancies: their category pages are what engines retrieve for best-of questions. A complete profile with recent, detailed client reviews on the two or three platforms your engines actually cite is the highest-leverage away-game move available, and you find out which platforms those are by reading the citations on the questions you currently lose.

Engines also disagree with each other, only about 11 percent of cited domains overlap between ChatGPT and Perplexity in cross-platform analyses, so the source map is per engine. The mechanics of that fragmented battlefield, and why it differs from rank tracking, are laid out in GEO vs SEO.

The Princeton GEO study found that statistics, quotations and cited sources lift visibility in generative answers by up to 40 percent, while keyword stuffing lands below baseline. For a services firm, the statistics are your case studies, and most firms write them wrong for this purpose. The GEO-ready case study leads with the number: pipeline grew 3.1x in two quarters, the audit closed in six weeks, the placement stuck for three years. It names the client where permission exists, or describes them precisely where it does not. It states the methodology in enough detail to be credible and the timeframe in enough detail to be checkable, and it includes a quotation from a named client contact.

Around the case studies, publish the niche-defining pages your buyers would ask an assistant about: what a HIPAA readiness engagement costs and how long it takes, what a fractional CFO actually does at $5M revenue, how to evaluate a demand gen agency before signing. One question per page, answered completely in the first 120 words, with an FAQ block and FAQPage schema. These pages win the exploratory questions from Semrush's data, the 59 percent using AI to understand a problem more deeply, which is where services relationships start.

The third asset is original research. Firms sit on anonymizable data no one else has: benchmark costs, timelines, failure rates in their niche. Publishing it creates the statistics other publications cite, and being the cited source in the pages engines retrieve is visibility that compounds. It is the same mechanism that makes review sites powerful, redirected through your own expertise.

Services expertise is attributed to humans, and engines follow that attribution. A boutique firm rarely outranks a directory, but a named partner can outrank almost anyone on a narrow question, because bylines, podcast appearances, conference talks and quoted commentary in trade press create exactly the third-party evidence trail engines look for when deciding who is credible on a topic. The practical program is steady rather than heroic: one substantial byline or quoted appearance per practitioner per quarter, in the publications your citations show the engines already reading, each anchored to the niche you want to own.

Public community participation belongs in the same program. The threads where founders ask which agency to hire, on Reddit and in professional communities, get cited by engines and read by buyers for years. Answering those questions substantively, under your real name with your affiliation disclosed, builds the recommendation corpus a services firm cannot buy. G2's buyer research found 85 percent of buyers think more highly of a vendor an AI recommends; the raw material of those recommendations, for services, is largely what named humans said about you and what your named humans said in public.

One mechanical note: put the expertise where crawlers can reach it. Podcast insights trapped in audio, talks that exist only as video, and posts locked inside feeds engines index poorly all evaporate for retrieval purposes. The fix is repurposing with intent: transcripts and written recaps on your own domain, author pages that tie each practitioner's bylines together, and Person schema linking the human to the firm. The expertise already exists; GEO for services is substantially the discipline of writing it down where machines read.

Compressed into one month for a ten-person agency. Week one: build the question list from your last twenty discovery calls, the phrasing prospects actually used, qualified by niche and stage; run the baseline across engines and store the answers. The same week, read the citations on every lost question and list which directories, rankings and publications keep appearing. Week two: evidence pass. Pick your three strongest engagements, secure naming permission for at least one, rewrite all three case studies to lead with the number and carry a named quote, and publish them as crawlable pages rather than PDFs.

Week three: presence pass. Complete the one directory profile your citations named most, request reviews there from the five clients who would write specifics, and pitch one byline anchored to your niche to a publication that appeared in the citation trail. Week four: publish one niche-defining page for the most valuable lost question, answer-first with your own engagement data as the statistics, index it in Google and Bing, then recheck the full question list against the baseline. One month, five assets, every hour traceable to a stored answer. Firms that repeat this loop quarterly tend to find the third iteration easier than the first, because the evidence pipeline starts feeding itself.

The most expensive mistake is positioning too broadly on the theory that a wider net catches more fish. Engines composing an answer for "demand gen agency for Series A SaaS" prefer the firm whose entire public record says exactly that over the full-service agency whose site says everything for everyone. Generalist positioning is a choice a firm can make for sales reasons, but it should be made knowing the cost: in generative answers, the specialist gets named and the generalist gets summarized into the phrase "among others."

The second mistake is confidentiality as a reflex rather than a policy. Plenty of client work genuinely cannot be named, but most firms never ask, and a case study with a named client, a real number and a quoted contact is worth ten anonymous vignettes to an engine weighing whether you are checkable. Build the permission request into your project close-out, when goodwill peaks, and the evidence pipeline fills itself.

The third is publishing thought leadership with nothing quotable in it. Essays about trends, unanchored by a number, a named source or a checkable claim, give engines no reason to cite you over anyone else saying the same thing. The Princeton findings apply to services content with full force: the post that says onboarding projects in your niche ran a median of eleven weeks across forty engagements will be cited for years; the post that says onboarding is important will never be cited at all.

The measurement discipline transfers unchanged from other verticals, with one services twist: your question list must carry the qualifiers. Fifteen to twenty-five unbranded questions with niche, stage and geography attached, run repeatedly across engines on a schedule, because SparkToro measured under a 1 percent chance that two identical ChatGPT runs return the same brand list. Store the raw answers, track mention rate per question and per engine, and keep branded questions out of the headline score. The citations behind each lost question are your work queue: they tell you whether the gap is a directory profile, a missing niche page or an absent evidence trail.

Then run the loop monthly. Fix the single highest-leverage gap per lost question, get new pages indexed in Google and Bing, and recheck against the baseline. Reachroller automates the cycle: scheduled runs across engines, honest mention rates with per-question receipts, the citation list behind every answer, and a publish-ready fix page generated for questions you lose, with slug, metadata and schema included. Starter is $29 per month for 400 credits and 25 tracked questions, sized for a firm rather than an enterprise martech stack; how it compares to the rest of the field is covered honestly in the best AI visibility tools. The caveat we always attach: it is a young product, ChatGPT tracking is live today and the other engines are rolling out. Start with the free check, three days and 50 credits, and see which firms the assistants name when your exact client asks your exact question.