A client asked whether they show up in ChatGPT. So you typed your own pitch into it, "best [your niche] agency," and watched it name three other firms. None of them yours.
That sting has a name now. "AI doesn't recommend my agency" is the quiet worry sitting under a lot of pitches this year, and it is more fixable than it feels.
Here is the good news. Engines do not invent the agency market. They re-rank a small set of sources they already trust, then assemble a list. Breaking into best agency AI answers is a supply problem, not a popularity contest.
This is the agency version of the play. By the end you will have eight steps you can run for every client, every quarter, as a productized agency AEO service line instead of a one-off experiment.
One thing before we start. The general diagnostic for any brand, mapping commercial prompts and baselining who wins them, holds here too. This guide inherits that backbone and layers the agency-specific levers on top: service directories, case studies, and niche authority.
Map the prompts your clients' buyers actually type
Start with the questions a prospect asks before they hire anyone in that niche. For each client, list 15 to 30 of them.
They cluster in tiers. Awareness prompts ("what does a [niche] agency do"). Consideration prompts ("best [niche] agency," "top [service] firms for [vertical]"). Comparison prompts ("[agency A] vs [agency B]," "big agency vs boutique"). And capability prompts ("agencies with case studies in [adjacent skill]").
The consideration tier is where the money is. Those are the exact queries that produce top firm AI recommendations, and they are the ones most agencies never track.
You are done when you have a flat tracker with one row per prompt, per client. Each row carries the buyer stage, the competitor firms you are watching, the client workspace, and which engines you monitor.
Where people go wrong: they let the client write the prompt list. The client asks "are we showing up," so the agency drafts brand-awareness prompts and skips the ranking questions entirely. Own prompt definition during onboarding. It is your job, not theirs.
This is one place DeepSmith saves you the blank page. Discover Prompts generates a starter set from each client's product, persona, and buyer-stage context, and the Prompts view then tracks every one of them with per-prompt mention and citation rates.
Baseline where you stand today, mentions and citations
Now run each prompt through every engine you cover and write down three things: whether the firm gets named, which firms do get named, and which exact sources the engine cited.
Turn that into three numbers per client. Mention rate is how often the firm appears at all. Citation rate is how often a page the firm controls is the source behind the answer. Share of voice is the firm's slice of all brand mentions across the query set.
Feeling behind already? There is more room here than you might think. Google AI Overviews typically pull from 8 to 13 sources per answer, spread across roughly ten unique domains. That is ten doors per prompt, not one.
You are done when each client has a baseline showing the share of prompts that return them at all, the share that cite a page they own, which competitors show up most, and which directories or roundups the engines lean on for their main consideration prompts.
Where people go wrong: auditing mentions and ignoring citations. Getting named through someone else's listicle is nice. It is also borrowed. An agency that never gets cited from its own page is one algorithm shift away from vanishing. Track both.
Pro tip: engines blend their trust signals. A cited directory profile, plus earned media, plus a clean page on your own site is the combination that earns you a mention and a citation inside the same answer.
DeepSmith reports these together. The AI visibility Overview gives you mention rate, citation rate, and share of voice with a per-platform breakdown and a competitor leaderboard, and Competitor citations shows the exact pages winning your prompts, engine by engine.
Fill out the directories AI actually cites
Service directories are the single biggest agency-specific lever, and most firms are half-present on them. When did you last open your own Clutch profile?
In a July 2026 audit of 50 vertical buyer prompts run through ChatGPT, directories showed up as cited sources in roughly 42% of the answers. Clutch dominated that set, appearing far more than any other. DesignRush, GoodFirms, and Semrush Partners followed, with The Manifest, Sortlist, G2, TopDevelopers, UpCity, the HubSpot Solutions Directory, Credo, and the Google Partners Directory filling the next tier.
For each profile, populate five things. A short positioning statement in the client's own wording, so an engine can lift it verbatim. The full service list, named the way buyers search ("social media marketing for B2B SaaS," not "social"). Five to eight case studies with real numbers and a named industry. Reviews collected through the directory's own verification process. Awards and ranking placements, displayed.
Clutch rewards specific behavior, so build for it. Rank there is driven most by the Ability to Deliver score, and reviews carry the heaviest weight inside it. Premier Verified status asks for an average rating around 3.0 or higher, at least three verified reviews, and at least one of those in the past twelve months. The Manifest sorts mainly on service lines, focus areas, and client reviews.
You are done when every profile is live, enriched, and carrying current reviews, and when you re-run the client's prompts a quarter later and see the firm turning up more often.
Where people go wrong: doing one directory well and stopping. Citation share rises with breadth, not depth on a single platform. The other mistake is filling profile descriptions with marketing fluff. Engines extract sharp, specific, stat-bearing sentences. Vagueness does not get lifted.
Rewrite your case studies so AI can lift them
Your case studies are probably your best proof and your worst-formatted asset. AI extracts sentences, not stories. The work is already done, so this step is reformatting, not rewriting.
Rebuild each one in seven parts. A one-sentence result claim at the very top, stat-bearing and self-contained, in the shape of "we cut cost per lead from $312 to $78 for a B2B SaaS client in 11 weeks." The client name and industry, stated plainly. The service line in buyer language. Two to four quantitative metrics, each with a baseline, a result, and a timeframe. A plain methodology line. A short quote from the client contact, with their name and title. An internal link to the matching service page.
Density matters too. Research shared in practitioner communities suggests pages carrying five or more statistics per thousand words, and sentences built to be quoted, see meaningfully higher citation rates. Treat that as a target to write toward, not a promise.
You are done when you can paste a buyer's actual prompt language into ChatGPT, Perplexity, and Gemini and see your result sentence surface. If it does not, rewrite the sentence, not the whole page.
Common mistake: writing the case study as a narrative essay about the journey you took together. Nothing in it is liftable. The second mistake is anonymizing the client for confidentiality. An unnamed claim is a weak claim to a model, so ask for permission and credit instead.
There is a reason this matters more than it used to. Forrester's 2025 B2B buying research found AI tools were the most-cited interaction in the buying process, ahead of contact with vendor salespeople. Your case studies are not just sales collateral anymore. They are the raw material of somebody's shortlist.
Producing these at roster scale is where DeepSmith carries weight. Content Studio turns a planned idea into a publish-ready article with structure, internal links, and metadata already built in, and Deep IQ holds each client's voice and product facts separately, so nothing bleeds between accounts.
Earn the niche authority signals that travel
Here is the number that reframes the whole job. Analysis of 25 million earned-media links found roughly 84% of AI citations trace back to earned media: original reporting, trade press, expert commentary, Q&A. Most of what wins the answer is not on your website.
Roundups are the other big surface. A Search Engine Land analysis of 25,000 cited URLs, covering around 400 million citations, found about 63% pointed to listicles. Which means being included, editorially, in "top ten [niche] agencies" posts is not vanity. It is citation supply.
Chase three things in order. Original reporting or data your niche's trade press will cover. Bylined leadership pieces in tier-one trade outlets with a full author bio. Recognition in real industry awards and third-party verifications, listed somewhere crawlable.
Outlet choice matters. Trade-press mapping in 2026 puts Ad Age and Adweek in the top authority tier for consumer and retail work, Forrester and Gartner for B2B technology, and vertical trades like PRWeek or MM+M for their own worlds. Titles a rung below, the ones easier to land, still get cited. Start there while you work on the tier above.
Wikipedia deserves a word, because it is the biggest single source in the mix. Reporting from May 2026 put Wikipedia at roughly 47.9% of ChatGPT's top-cited factual sources, and earlier platform studies put its share lower, in the single digits to low teens depending on the month. Either way it is large. For most mid-market firms a Wikipedia article is still the wrong target, since organizational notability demands significant coverage in multiple independent secondary sources and rules out press releases and directory listings. Aim instead to be a sourced reference inside an existing category article.
You are done when the firm lands at least three earned mentions a quarter and has a press page in plain HTML, listing publication, date, headline, and link.
Where people go wrong: buying placements and sponsoring listicle slots. Engines weight editorial credibility, and paid inclusion travels badly into citation sets. Press relations, bylines, and original research are the durable plays.
Mark up your site so engines can read it
If you searched for how to get agency cited by AI and landed on a schema tutorial, this is the part people skip. Do not skip it. JSON-LD is the least ambiguous way to tell an engine what your pages are.
Cover the core types. Organization on the homepage, with name, logo, contact point, founding date, and a sameAs block linking your LinkedIn, Crunchbase, and directory profiles. That sameAs block does the most work, because it is how engines confirm you are one entity across the web. Then Service schema on each service page, Article or CreativeWork on each case study, FAQPage wherever you have a question block, HowTo on process pages, and BreadcrumbList site-wide. Speakable shows little citation effect so far, so treat it as optional.
FAQ markup is worth the effort even on light pages. In Ahrefs research across 863,000 AI-mode result pages, FAQ schema appeared on around three quarters of top-ranking pages and HowTo on roughly a quarter.
That same research carries the most encouraging finding in this whole guide. As of its March 2026 update, only 38% of AI Overview citations came from the top ten organic results, down from about 76% a year earlier. You do not have to outrank the giant firm to be quoted next to it.
You are done when Organization, Service, and Article markup pass the Google Rich Results Test and the Schema.org validator with zero errors, and you have a quarterly reminder to re-check.
Where people go wrong: marking up one page and calling it done. Engines trust consistent markup across a site more than a single decorated page. Schema also breaks quietly after CMS updates, which is exactly why the quarterly re-validation exists.
Refresh every quarter with new numbers, not new words
Publishing is the start of the clock, not the end of it.
Every quarter, go back through the client's case studies and service pages. Update the metrics. Show a visible "last updated" date. Add a line explaining what changed. Freshness is a soft signal, and in a crowded answer set soft signals decide things.
Page quality compounds here. In the GEO-16 testing, pages that cleared an upper quality bracket, a composite of citations earned, freshness, structured-data correctness, and topical relevance, held citations across engines at a much higher rate than pages below it. Consistency across engines is the payoff for doing the boring maintenance.
You are done when every case study and service page carries a published date and an updated date, and each quarter produces at least one new case study plus one or two refreshed service pages.
Where people go wrong: treating a refresh as a copy edit. New wording moves nothing. New data moves a lot. Never ship a refresh without a number in it.
Report it back in a client-ready dashboard
None of this counts until a client can see it in a report with your logo on it.
Give each account a standing dashboard covering five things: per-prompt mention rate, per-prompt citation rate, page-level attribution showing which pages earn citations, the competitor leaderboard, and the top cited sources per engine. Send it monthly. That cadence is what turns agency AEO from a project into a retainer line.
Which engines you can report on depends on the plan you run. DeepSmith Pro is $99 a month, or $80 billed annually, and covers ChatGPT. Grow is $199, or $160 annually, and adds Perplexity. Scale is $399, or $299 annually, and adds Gemini. Enterprise is custom and covers all five engines, including Claude and Google AI Mode, with 1:1 onboarding and a dedicated account manager. There is a 7-day free trial with real data and real drafts, and no long-term contracts.
You are done when every client receives a monthly white-labelable report showing citation wins, share of voice against their named competitor set, and the placements earned that quarter.
Where people go wrong: running the whole roster out of one workspace. Each client needs isolated context, its own tracked prompts, and its own competitor set. Cross-client bleed is the failure mode to design against, and it is much easier to prevent than to explain.
What to do next
If eight steps feels like a lot, that is normal. You are not doing them all this month.
Pick one client and one step. For most agencies the fastest movement comes from step three, because directory profiles are usually half-filled and the fix is measured in hours. Complete Clutch first, then work down the list.
Expect the timeline to be honest with you. Most firms see measurable movement within 60 to 90 days of a proper directory and case-study pass. Real share-of-voice gains usually take two or three quarters of steady earned media on top. That is the same answer you should give clients, by the way. It builds more trust than a promise you cannot keep.
You do not need a bigger delivery team to run this. You need one repeatable system across every account. If you want to see where each client actually stands and start closing those gaps with on-brand pages, you can start a DeepSmith free trial and work from live data on day one.
