How marketing agencies get cited in ChatGPT, Perplexity and Google's AI Overviews - the 2026 playbook

Most agency-side thinking about AI search is about the client. This piece is not about the client. This is about how your agency itself gets named when a marketing director opens ChatGPT, Perplexity or a Google result page and asks who they should hire.

The mechanics are different from ranking on Google, and they are different from getting into trade press. They overlap with both, and the overlap is where the returns compound. The shape of the work is its own thing, and the agencies that treat it as its own thing are being cited three quarters into 2026. The agencies that treat it as "SEO with a different name" are not.

The numbers make the case briefly. Google's AI Overviews now surface on 48% of queries and 58% of searches end without a click, per Search Engine Journal's June round-up of the data. Brands cited inside those AI answers see roughly 35% more organic and 91% more paid clicks against a baseline. When a marketing director types "best B2B PR agencies UK 2026" today, half the time they read the AI answer and never scroll. The agencies inside that answer are the shortlist.

The unit of visibility is no longer the ranking. It is the citation.

Here is the playbook we run for the agencies we work with, and use ourselves.

Start with entity structuring, not content

Language models pick names to cite based on which entities they can resolve unambiguously. If your agency has three subtly different names across your website, LinkedIn, Companies House and trade-press mentions, the model treats you as three weak signals rather than one strong one, and cites someone else.

The first job is boring and cheap. Reconcile the agency name, the founder names, the office locations and the client list across every public surface. Match Companies House filings to LinkedIn to the website footer to Crunchbase to Wikipedia. Where an old brand name still lives on an archived press release, decide whether to redirect it or leave it - both are defensible, but pick one and be consistent. The entity audit piece walks through the specific checks.

An hour on this beats a month of writing.

Own a small, real category

Models do not cite "great UK agencies". They cite named winners of resolvable categories. "Best independent B2B PR agency for fintech in the UK" is a resolvable category. "Best agency" is not.

Pick two or three narrow categories where you can honestly claim first, second or third position. Write the category page on your own site with the criteria a marketing director would use, not the criteria you wish they would use. Populate it with named clients who match. Repeat the same category framing across trade-press quotes, your LinkedIn, your MD's speaking bios and any award entries.

Six months of consistency in one narrow category will beat two years of hedging across five broad ones. The reason is that models score coherence heavily, and hedged category claims read to a model as low-confidence data.

Get named in the training corpus, not just indexed

There is a difference between being indexed by Google and being inside the training data that LLMs pull from. The models that matter today - GPT-5.x, Claude, Gemini, Perplexity's routers - pull heavily from a specific set of surfaces:

1. Editorial trade press with named bylines, especially Campaign, PRWeek, The Drum, Marketing Week, Adweek and Little Black Book.
2. LinkedIn posts and comments from accounts with an established audience, particularly founder and MD accounts.
3. Podcast transcripts that have been mirrored to sites like Rev, Descript or the podcast's own show notes.
4. Wikipedia and Wikidata entries and the databases they seed - Crunchbase, PitchBook, Companies House, Google Knowledge Graph.
5. Community threads on Reddit, Hacker News and specialist Slack digests where they get archived to the open web.

You want a presence in at least three of those five. One is a signal. Three is a pattern the model can index.

The LinkedIn cadence piece covers what "presence" means on LinkedIn specifically. The Campaign timeline piece covers the trade-press side.

Structure the site the way the model reads it

Language models do not read your site the way a user does. They tokenise it and score which passages are most likely to answer a query. The site changes that matter most:

1. One page per resolvable question. If you want to be cited as the answer to "how do independent agencies price AI work in the UK", there should be a page on your site with that exact question in the H1.
2. A single-sentence direct answer in bold at the top of that page. The FAQ-schema pattern the model is trained on rewards this shape heavily.
3. Named-author bylines on every long-form post. Anonymous editorial gets scored as low-authority. Named partners with LinkedIn links are the shape of a citable source.
4. Structured data - Article schema with author, publisher, datePublished; FAQ schema on question pages; Organization schema with sameAs links to LinkedIn, Crunchbase and Wikipedia if you have one.
5. Clean internal linking. Every post that touches a category should link to the category page in natural sentence anchors, not "click here". The model uses the anchor text as part of the entity signal.

None of the five are expensive. All five compound.

Feed the loop with weekly signal

Citations are not a one-shot exercise. They are a rolling index. Every week you go quiet, another agency writes the trade-press piece that is going to be scraped, and the model's next refresh nudges toward that agency.

The realistic weekly minimum for a mid-sized indie is: one MD LinkedIn post with named specifics, one expert-comment landing in a trade publication, one on-site publication of your own with a resolvable H1. That is three surfaces a week. It is not exotic. Most agencies do all three already, but not consistently. Consistency is the whole game.

The job-title piece describes how to structure your named team's public presence so the model can match roles to categories. Do that once, and every future weekly post inherits the entity structure.

Measure what you can, ignore what you cannot

You cannot see inside a language model's training weights. You can see the outputs. Four measurements are useful:

1. Direct query testing. Every fortnight, run your priority category queries in ChatGPT, Perplexity, Claude and Google AI Overviews. Note when your agency is named, when a competitor is, and what the citation text says. Track it in a spreadsheet.
2. Referral traffic tagged as ai in analytics. Perplexity, ChatGPT and Copilot pass identifiable referrers. Watch the trend, not the absolute number.
3. Trade-press mention rate against a competitor set. If competitors are getting quoted in Campaign twice a month and you are getting quoted twice a quarter, close the gap before you optimise anything else.
4. Named LinkedIn mentions of the agency and its people. The comment volume and named-tag rate are what the model sees as social proof.

If those four are moving in the right direction over a quarter, the citation rate is following. If they are flat, the model is not going to invent a reason to name you.

The four traps agencies fall into

Most agencies that start this work stall in one of four predictable places. Naming them upfront saves months.

1. Writing content before fixing the entity. The site publishes weekly, the model has nothing structural to attach the content to, and none of it compounds. Do the entity reconciliation first, always.
2. Chasing broad keywords instead of narrow categories. "Best marketing agencies" is a query the model answers with WPP, Publicis and IPG. It will never answer it with your agency. Pick something the model can actually award you.
3. Publishing on the site but not the trade press. The model reads your site as owned media. It reads trade press as third-party validation. One without the other is half the work.
4. Optimising for one model and forgetting the others. The prompts and citation formats that ChatGPT rewards are different from what Perplexity rewards, and both are different again from Google AI Overviews. If you only test one, you will only be cited in one.

None of the four is fatal. All of them cost a quarter.

The build order

The full playbook is a six-month build. The order matters more than the pace.

1. Weeks one to two - entity reconciliation and structured data on the site.
2. Weeks three to six - one category page live, one MD post a week, one trade-press pitch a week.
3. Weeks seven to twelve - second category page live, expert-comment cadence established, first Wikipedia or Crunchbase updates land.
4. Weeks thirteen to twenty-six - measure, adjust, publish weekly, expand into a second category or a second named leader.

At the end of that, you are inside a rising share of the answers, and the compounding starts. Agencies that skip the first two weeks and jump to publishing find the writing does not stick - the model has nothing structural to attach it to.

If you want a starting point mapped to your own agency, the method page walks through how we run it with clients. If you want a conversation about which categories you could realistically own, get in touch and we will trade notes for an hour.