Expert Strategies
for B2B SaaS Marketing

Short Version

In early 2025 we ran 100+ client websites with 30 people and a bench of more than 100 freelance writers. A blog post took four days. We could publish 10 to 12 pieces per client per month, and that was about our limit.

Today the same post takes around two hours and costs about $8. We publish 25 to 30 pieces per client per month and refresh 100 more. Of course, one strategist goes through everything before it is published.

We got there by building our own platform, Slate, and moving our organic delivery onto it. How did we do it?

1. What Our SEO Retainer Looked Like in Early 2025

About 80% of the team worked on SEO and content, and 20% on web development. At our peak we had 14 in-house writers and 4 editors on top of the freelance bench, and five roles touched every account:

•  A strategist owned the plan and the calendar

•  Writers drafted

•  A technical SEO person handled crawl, schema, and site health

•  An account manager held the client relationship and built the monthly deck

•  Web developers did anything that touched the site

Every Deliverable Was Priced in Human Days

‍

Most of the Hours Went to Planning and Audits

The two things clients cared about most, publishing and refreshing, were the fastest things we did. The slow items were audits, planning, and reporting.

•  Competitor research redone every quarter, for every client

•  Audits that produced a document but no change to the site

•  Implementation queues waiting on a developer

•  Reporting at one working day per client, which across 100+ sites is 100+ person-days a month, or about five full-time people assembling slides about work that had already happened

What Was Good About This Model

We didn't want to lose any of this:

•  A senior person saw every page before it went live

•  Strategy and execution sat in the same head, so a strategist who spotted something odd in the SERP could act that afternoon without filing a ticket

•  Clients knew the names of the people doing their work

We kept the first two on purpose, and chapter 8 covers how.

On a $10,000 monthly retainer the money went roughly like this: $4,500 to writers and editors, $1,800 to account management, $1,500 to strategy leadership, $400 to tools. That left about 18%.

2. What Happened When We Tried to Grow

Content Volume Was the First Constraint

Ten to twelve pieces a month per client is fine until the client wants forty, or until the plan calls for a few hundred programmatic pages built off a template. At that point the only honest answer was that we'd need more writers, and more writers made everything below worse.

The Same Brief Gave Us Different Articles Depending on Who Picked It Up

With more than 100 freelancers in rotation, two writers working from one brief handed back two different articles, at different reading levels, with different takes on the client's voice.

We handled it the way most agencies do, by routing the accounts that mattered most to the writers we trusted. That works until the trusted writers are full. After that the routing decision is doing the work of a quality decision, and nobody set out to make it that way.

Our Best Methods Lived in Two People's Heads

Two of our strategists were better than our documentation. One understood SERP intent better than anything we'd written down. Another had a way of diagnosing a page for refresh that beat the generic checklist.

It never existed in a form someone else could pick up and run. Our playbooks said what to do, not how the good operators decided, so nothing compounded and what we learned on account 40 stayed on account 40.

Training a new hire meant a quarter of shadowing. We were teaching taste, one person at a time.

Every New Client Needed Another Slice of Three People

Add a retainer, add roughly a proportional slice of a writer, an AM, and a strategist. Revenue went up and margin stayed where it was. At 18% there isn't much room to absorb a bad quarter or fund anything that doesn't bill.

3. Where All of That Was Coming From

It Took Four or Five Tools to Finish One Page

We took a long time to notice this one.

Each tool was good at the step it owned. None of them handed off to the next, so the operator did the joining up.

We never counted that as work, because it looked like using tools. As far as we can tell it was the biggest single cost in the process, which meant scaling content mostly meant scaling the person in the middle.

We used a lot of tools to write content, whether it be ChatGPT, Perplexity or Claude. Then for keyword research we found another tool. It was a lot of tools all over the place, and we couldn't really scale it, because we were using different tools to get one single use case completed.

— Labeeb Muhammed, GTM Engineer, TripleDart

We Had 100+ Accounts of Evidence and No Way to Use It

We ran organic across fintech, dev tools, HR tech, CX, and travel, and every account was learning something about its category that never pooled. Ask us in early 2025 which heading structure won featured snippets across the portfolio and the honest answer was a strategist's instinct.

Brand guidelines lived in Google Docs, one per client. Competitor lists sat in spreadsheets. Approved statistics sat nowhere, so every writer researched the same numbers again.

Buyers Started Asking ChatGPT, and We Couldn't Measure That

Through 2025 our clients' buyers started asking ChatGPT and Perplexity the questions they used to type into Google, which is a different optimisation problem from ranking on ChatGPT. Our measurement, and pretty much everyone's, still watched blue links.

We could tell a client their position on a keyword, but not whether an AI answer to the same question mentioned them, cited them, or cited a competitor.

Over the last year, I've had more conversations about ChatGPT, AI Overviews, Reddit, and LLM visibility than traditional Google rankings. The way people discover software is changing. SEO is no longer just about earning blue links. It's about showing up wherever buyers are researching, comparing, and making decisions.

— Glenn Gabriel Bona, Associate Director, Organic Growth and Community, TripleDart

Our First AI Drafts Needed More Checking Than Writing

We'd been using general-purpose models for drafting since 2024 and they worked up to a point. Past that point they came back with made-up statistics, citations to sources that didn't exist, internal links quietly dropped, and not much of the client's voice.

You can still see the effect in our workflow list: a Fact Checker, an AI Fact Checker, and a reading-ease and citation-worthiness analyzer all sit there as separate live workflows, each one built because output failed in that exact way.

Everyone has AI. So why does most AI content feel generic? Because the bottleneck was never writing. It's knowing what to say, who you're saying it to, and how to make someone actually care.

— Manoj Palanikumar, Co-founder and Head of SEO and Content, TripleDart

4. What Made Us Change

It Started With a Programmatic SEO Build

We set out to run a programmatic SEO build, generating a large set of pages against a template and a data source. That's an ordinary thing for an SEO team to want.

We couldn't get it to the volume the plan needed, and the writing wasn't the hard part. Moving one page through five disconnected steps was manageable. Multiplying that by a few hundred was a staffing request.

The Question We Kept Coming Back To

That's where we thought about having a tool of our own. Why can't we have a tool on our own? Different functionalities, different use cases, in one single tool. From research to your analytics, your insights, opportunities, to content.

— Labeeb Muhammed, GTM Engineer, TripleDart

Not one tool for writing, which we already had, or one for keyword research, which we also had. One place where each step's output arrived as the next step's input, without a person carrying it across.

The measurement gap in chapter 3 was pushing in the same direction, and it wasn't really an AI problem. We'd built a delivery business on one metric, position, that a third party handed us for free. Twelve years of running organic and we owned none of the measurement.

Then We Noticed the Numbers Wouldn't Sit Still

Our first instinct was to treat prompts like keywords: build a list, check the brand's position in each answer, report a share-of-voice number. It was the same operating model pointed at a new surface, and you can see that thinking in our oldest workflows.

Then we ran the same prompt twice on the same day and the cited sources moved. Run it on a different engine and the set of cited domains barely overlaps.

A ranking position is stable enough to report monthly. A citation isn't. So we had to change how often we measured before the metric meant much, and continuous measurement isn't something you do by hand.

5. How We Automated

First Go: Three Fixed Tools

We picked the three jobs eating the most hours and built a tool for each: an AI content writer, a content optimizer, an internal linking identifier.

All three worked in a demo. In production they hit the same wall: one fixed template couldn't serve 100+ different websites. The writer tuned for a dev tools client gave us something unusable for a fintech client, and serving both meant adding configuration until we had a worse version of a general-purpose system.

What we took from it: we needed a canvas, not a fixed set of features.

Second Go: Every Operator Built Their Own

So we gave operators a canvas. Within weeks we had article writers, refresh workflows, and blog builders running in individual accounts, each built by the person who ran that account. Operators solving their own problems are generally faster than a central team guessing at them.

Then we counted. One client account had five separate workflows doing rank and position tracking, built by five people who each thought they were solving it for the first time. Four were used once. So we paid for that build five times.

Third Go: One Shared Library

So we consolidated into one versioned library any account can clone, with a brand kit per client that the workflows read from. Research, briefing, drafting, and refreshing came together quickly. Publishing did not.

Writing is a language problem, and language models are good at those. Putting an article into a CMS is a formatting problem across:

•  Rich text fields and image references

•  Slug rules and schema blocks

•  Internal link markup

•  A publish API that will happily accept your payload and mangle the body HTML

Every client CMS is set up differently, and each difference is another way for a publish to succeed and still be wrong.

Our Webflow Publisher Is on Version 20

Our three Webflow publishers carry 46 versions between them. Our brand-kit article writer is on version 1, and so is our content brief generator.

The writing half was close to right the first time. The publishing half took 46 goes.

Eight Things We Started and Stopped

‍

We Built the Checks Before We Turned Up the Volume

On refreshes there's one more pass: every statistic already on the page gets checked against the live web before anything is rewritten.

The Rule We Ended Up With

Checking has to cost less than writing.

Under the old model, writing cost far more than checking, so reading every word by eye made sense. Once a draft costs $8 and 12 minutes, the person reading it becomes the most expensive step in the run.

So the checks moved into the pipeline and the person moved to the end of it. A strategist now sees a finished, verified, publish-ready draft instead of a raw one.

6. Why We Built Instead of Bought

We bought for a couple of years before we built anything, and the tools we bought were good at their own step and never owned a job end to end. Four questions stayed unanswered:

1.  Across a hundred buyer questions, how often does an AI answer name this brand, and what does it cite instead?

2.  Which of this client's pages have decayed enough to need a refresh this week, without a person checking?

3.  What does this client's brand voice look like as structured data a machine can apply?

4.  Can a research step feed a writing step feed a publish step, with nobody moving a file in between?

What Running 100+ Accounts Let Us See

Most software companies building for marketers have to go and find out what marketers do. We had 100+ live accounts generating that evidence daily, and a delivery team who'd tell us within the hour when something didn't work.

An agency portfolio is a running experiment across categories, so one measurement layer across all of them shows which patterns hold everywhere and which are local to one account. We'd expect more of this kind of software to come out of agencies than out of vendors.

7. How We Built It, and What the Data Showed Us

Measurement First, Writing Second, Publishing Last

We built in this order, over about ten weeks:

1.  Trackers. SERP intent, share of voice, and which competitors we weren't watching yet.

2.  A quality scorer. Reading level, and whether a passage is quotable by an AI engine.

3.  Writers. Article and brief generation against a brand kit.

4.  Refresh. Audit, diagnose, rewrite.

5.  Publishers. Into the CMS.

Of everything here, the sequencing is probably the most portable.

Each brand now gets a prompt set running continuously across ChatGPT, Perplexity, Gemini, Claude, AI Overviews and AI Mode, grouped into topics so the analysis rolls up to a whole category. The GEO metrics we report on come out of this layer.

How We Sort Each Cited Source

We spend most of our time on "not mentioned", which the old model had no name for.

We call it the mention gap: the share of sources an AI engine cites on a brand's own topics that never mention that brand. A brand can be missing from an answer two ways. Either no answer got generated on the topic, or an answer got generated, cited five sources, and none of those five knew the brand existed. The second is a content gap with an address on it.

It's adjacent to what Kevin Indig calls a ghost citation, where an engine cites a page without naming the brand. The mention gap looks at the same problem from the brand's side and asks what share of the cited set it is missing from, which is the version an operator can act on. If you're new to this, our primer on generative engine optimization covers the underlying mechanics.

What We Found Across Four Client Accounts

Over the 90 days to 13 August 2026 we parsed 667,250 citations across four live client accounts. In all four, the biggest bucket was a cited source that never mentions the brand.

‍

Two things stand out:

•  The gap sat between 44% and 77% in all four accounts, across four unrelated categories

•  A brand's own pages account for 0.7% to 8.3% of citations, so almost nobody owns their own

Everstage leads at 8.3% and also carries the highest competitor share at 18.2%, which suggests a more contested category than the others.

One caveat on the floor of that band. TripleDart's 43.6% is the lowest figure in the table, and it is probably understated: nearly half of our own citations fall into an unclassified bucket rather than being sorted against a competitor set, which is a configuration difference in our workspace rather than a real difference in the market. Read 44% as a conservative floor. The three client accounts run from 52% to 77%.

A larger study run on the same tracking backs up the shape of this. Across 300,000+ citations for six B2B SaaS brands over 90 days, spanning ChatGPT, Perplexity, Gemini, Claude, AI Overviews and AI Mode, 76.3% of citations went to domains that were neither the brand nor a competitor. Average owned citation share came out at 3.3%, with the best performer at 8.6%. Our four-account band sits inside that, and the ceiling is almost identical.

We looked at 150,000+ AI citations in one B2B category over 90 days. Only 3.6% pointed back to the brand's own site. Where did the other 96.4% go? 33.7% to competitors, 30.9% to social, 31.8% to third-party pages that don't even mention you. In B2B, whoever writes those lists is basically the encyclopedia the AI reads from.

— Shiyam Sunder, Founder, TripleDart

A program that only publishes on the client's own domain is competing for the smallest slice on that chart. So citation gap outreach, aggregator sweeps, and Reddit responses now sit alongside the content workflows.

The Twelve Jobs We Automated First

These run unattended on every account we manage. If you're working out where to start, this is roughly the order we'd go in now.

‍

8. What Changed in How We Work

The Day Stopped Being a Circuit of Platforms

An SEO operator now opens one workspace in the morning and stays in it. Research, tracking, briefs, drafts, refreshes, and publishing all happen in the same place.

I used to spend the first hour of every day pulling numbers before I could make a single decision. Now they're already there when I open the workspace, so that hour goes on deciding what to do about them instead. That's the part of the job I wanted anyway.

— Anand Raju, SEO Lead, TripleDart

We Stopped Opening Client Calls With Rankings

A page can hold position three and be missing from every AI answer on the same question. Another can get cited constantly while sitting on page two. We still track position, it's just not the number we open on.

We review accounts on this instead: of the AI answers generated on a brand's topics, what share cite a source that names the brand, and what share cite a source that doesn't. It's closer to answer engine optimization than to classic rank tracking. That single number gives you the size of the opportunity and roughly where it sits.

Reporting changed shape too. Clients get their updates out of the same system that did the work, so nobody assembles a deck afterwards.

Publishing More Helped Less Than We Assumed

Under the old model, output volume was the visible sign a retainer was working, so the incentive was to publish.

An account publishing 30 new pages a month and refreshing nothing is building a library that goes invisible on a rolling 13-week cycle. We were doing exactly that on several accounts without realising it.

So our refresh volume went from 20 pages a month to 100, and the top 200 pages on an account now get audited weekly.

Most SEO teams are sitting on a goldmine of traffic they're barely maintaining. The majority of your organic traffic probably comes from old content, but in 2026, refreshing content isn't just about rankings anymore. It's about: preventing traffic decay, winning AI citations, keeping content accurate at scale, building workflows that don't require manual updates page-by-page.

— Glenn Gabriel Bona, Associate Director, Organic Growth and Community, TripleDart

We Started Working on Pages That Aren't the Client's

Listicle displacement, aggregator sweeps, and Reddit threads are part of the organic program now, sitting in the same workflow library as the content jobs because they do the same work on a different surface.

We came to this late. We spent years getting good at producing pages and only months getting good at appearing on other people's.

Three Job Titles Got Rewritten

The freelance writing bench went from more than 100 writers to 10 reviewers. The people we kept are the ones who edit and decide.

The role is no longer about rankings, traffic reports, and long roadmap decks. I'm looking for operators who understand AI visibility, brand mentions, citations, and GEO metrics.

— Manoj Palanikumar, Co-founder and Head of SEO and Content, TripleDart

Four Things We Kept Human on Purpose

1.  Strategy. What to publish, which topics to contest, what to leave alone.

2.  The last look before publish. One strategist goes through the finished draft. Every published page still passes a person.

3.  Client relationships. Nobody gets an automated account review.

4.  Anything written in a named person's voice. It's why we stopped automating HARO replies. Finding the opportunity is fine. Writing a quote attributed to a human being isn't something we're comfortable handing over.

9. What a Week Looks Like Now

Most of the configuration sits in the brand kit. It holds a client's voice, guidelines, competitors, and personas as structured data, and every workflow reads from it, so the same pipeline produces different output for a fintech client than for a dev tools one.

What Runs on Monday Morning

The rank tracker pulls positions, and pages showing decline get classified and routed into the refresh queue with a diagnosis attached. The prompt monitor runs each brand's prompt set across the engines and updates the citation picture. The mention scraper sweeps the aggregators, and the competitor tracker logs what moved.

None of this needs anyone to start it.

A New Blog Takes About 12 Minutes to Draft

Keyword and brand kit go in. Ahrefs pulls SERP and keyword data, competitor pages get scraped and their structures extracted, and an outline gets built against the SERP patterns and the brand's angle. Deep research assembles verified sources, the draft gets written against the brand kit, and internal links get mapped off the live sitemap.

‍

Laid out end to end, the logic looks like this. The decision point in the second column is where a person can pick competitor pages, or let the workflow take the top results and carry on.

‍

A Refresh Takes About 9 Minutes

Diagnosis is what makes the rest worth running. A page losing rank needs a different treatment from a page sitting just off page one, and before this step existed we treated them the same way.

Then a Person Goes Through It

A strategist works through the finished draft, which is the only step we've kept human. Then it publishes into the CMS with schema and internal links already in place.

•  Runs on its own: rank tracking, decay detection and routing, prompt monitoring, citation gap detection, mention scraping, competitor tracking, fact checking, linking, schema injection, report assembly

•  A person handles: the topic plan, the finished draft, anything in a named individual's voice, and every client recommendation

One Operator Now Runs Five or More Accounts

An SEO operator's week used to be execution and chasing. Now a single operator carries five or more clients without losing track of any of them, because most of what fell away was coordination.

I'm easily able to handle five plus clients without having to think about a lot of the execution going on around me. Slate focuses on the execution, and I focus on the strategic side of things. It also keeps me in the loop, so I know all the deliverables are up to the mark.

— Labeeb Muhammed, GTM Engineer, TripleDart

What Makes It Add Up Over Time

•  Workflows are reusable. Build one for a client, clone it, point it at another brand kit, run it.

•  Brand kits are permanent. Structured data doesn't decay the way a Doc does, and improves every time somebody corrects it.

•  Every run adds evidence. The citation record is what let us put real benchmark numbers in chapter 7. A year ago that section would have been one strategist's opinion.

Relationships and individual expertise used to be our main assets, and both left with the person. Workflows, brand kits, and the measurement record stay in the business.

10. The Numbers, Before and After

Q1 2025 baseline against Q1 2026 actuals. SEO, content, and web only.

What Each Deliverable Costs Now

‍

CRO audits and implementation weren't part of the Q1 2026 measurement set, so we've left them out here.

What It Did to the P&L

Look at the account management line. It stops being a separate cost once reports assemble themselves, because the strategist can hold the client relationship without losing a day a month to slides.

11. What It Produced

RateGain: 7 AI Overview Citations to 85, in 18 Days

RateGain sells software to travel and hospitality companies. Their organic rankings were strong and their presence in AI answers was not: across 484 tracked keywords they appeared in just 7 AI Overviews, and 278 of those keywords weren't ranking in the top 100 at all.

We didn't write anything new. We ran their existing pages through the refresh workflow, restructuring them into the question-and-answer shape AI Overviews pull from, with the direct answer in the first fold.

Eighteen days later:

‍

54 of the 85 citations were newly earned rather than converted from pages already ranking, and 23 came from keywords that hadn't ranked at all. The booking engine cluster, which was the priority, went from 0 to 28.

RateGain is also the second row in the table above, which is the same account measured across every engine rather than just AI Overviews.

Razorpay: Category Leadership, Measured Daily

A multi-business-unit fintech with hundreds of product pages and no single view of AI search across six platforms. Monitoring was manual, so a competitor comparison took weeks and was stale by the time it was finished.

We put continuous tracking across those platforms into one workspace, with workflows for citation gap analysis and comparison page generation on top.

Razorpay's own domain now generates 39,783 citations, and the most-cited sources are high-intent pages rather than blog posts: /pricing at 3,359 cited runs and /payment-gateway at 2,849. Their visibility share is close to double that of the nearest competitor.

The change worth naming isn't the number, it's that checking went from scattered manual passes to something that runs itself. The full write-up sits with our other customer case studies.

Our Own Brand, Through Our Own System

We run TripleDart on this too. 14 in-house writers and 4 editors producing about 30 blogs a month became a much smaller team producing five times the output, at 78% lower headcount, $8 a page, and 40% EBITDA margin across FY24 to FY26.

Our own owned-citation share is 0.7%, the lowest of the four brands in that table. We spent 2026 building the system for clients and our own citation surface got whatever was left. This report is the first bit of the correction.

The thing I didn't expect was how much less chasing there is. Nobody asks me where a draft is, or whether a page got refreshed. It either ran or it flagged, and I can see which one. What I bring now is the judgement about what's worth doing, and that used to get squeezed out by all the coordination.

— Siddu, SEO Lead, TripleDart

What We're Still Working On

•  Off-domain distribution. Most of the winnable ground is third-party pages, and those LLM SEO workflows are our newest and least mature.

•  Anything in a named person's voice. We stopped after the HARO responder and haven't found an approach we're happy with.

•  Publishing. Still the slowest part of onboarding a client on a CMS we haven't seen before.

A note on the limits: four accounts is a small sample, two are B2B SaaS, one a marketplace, one travel tech, and the lowest reading in the set is our own account, where classification is configured differently. Treat the band as a starting reference, not a market-wide figure.

Three Things You Could Try on Your Own Brand

None of these need our platform. They need an afternoon.

1. Measure your mention gap. Run 20 buyer questions from your category through ChatGPT and Perplexity, and mark whether each cited source names your brand. Ours landed between 44% and 77%. If yours is in that range, third-party coverage is probably worth more than another post on your own domain.

2. Check how old your cited pages are. About half of AI citations point at pages under 13 weeks old. If your most important pages haven't been touched in a year, decay is probably already running.

3. Count the tools it takes to finish one page. List every platform your last article passed through, from keyword to live URL, then count the handoffs. Ours came to five, and that number was what capped our volume.

Common Questions

What is the mention gap?

The share of sources an AI engine cites on a brand's own topics that never mention that brand. Across 667,250 citations in four B2B accounts we measured it between 44% and 77%, and between 52% and 77% across the three client accounts.

What is a good AI citation share for a B2B brand?

In our sample, a brand's own pages accounted for 0.7% to 8.3% of the citations it appeared in. A larger study across six B2B SaaS brands put the average at 3.3% and the best performer at 8.6%, so treat single digits as normal rather than as failure.

How do you measure whether AI answers mention your brand?

Run a fixed set of buyer questions across ChatGPT, Perplexity, and Gemini on a schedule, then classify every cited source as owned, mentioning, competitor, social, or not mentioning. A monthly manual sweep won't hold, because the same prompt returns different sources on the same day.

How often should content be refreshed for AI search?

About half of AI citations point at pages under 13 weeks old, so a quarterly cycle on the pages that matter is the floor. We audit the top 200 pages on an account weekly and refresh roughly 100 a month.

Does publishing more content improve AI visibility?

Not on its own. An account publishing 30 pages a month and refreshing none is building a library that ages out on a rolling 13-week cycle. Refresh volume moved our numbers more than new output did.

How many people does an AI-native SEO team need?

One operator now carries five or more accounts here, against 2.2 sites per head across the old team. The work that remains is topic strategy, the final read before publish, and the client relationship.

Is this the same as ghost citations?

Related. A ghost citation is an engine citing a page without naming the brand. The mention gap measures the same behaviour from the brand's side, as the share of the cited set it is absent from.