In this Mucker Growth Series session, we invited Irina Maltseva, founder of Seen, to share how early-stage B2B SaaS companies get found on Google and inside LLMs. She has spent close to ten years in SEO and GEO, first in-house, then advising, and now runs an agency doing search and AI visibility strategy, narrative content, and third-party mentions. A recap of what she covered, along with the full webinar recording and her slides, is below.
The Pattern Behind Invisible Content
After hundreds of conversations with startups, Irina sees the same three problems with content getting found.
- The content is something nobody asked for, because someone inside the company decided it was a good idea.
- The content is genuinely good, built on real customer research and written well, and still nobody sees it.
- Or the expectations were wrong from the start: publish today, expect traffic tomorrow, give up in month two, track the wrong metric, and have no way to attribute anything.
Her framing is that these are the same problem wearing three different outfits.
Search Didn’t Die. The Clicks Did
You have seen the LinkedIn posts declaring SEO dead and GEO the future. Her read is that this isn’t right, but it isn’t completely wrong either.
Google is competing with the LLMs by becoming more like them. AI Overviews answer the question directly. AI Mode is a different search experience entirely.
The effect on publishers is measurable. According to Ahrefs research, clicks to the number one result on Google fell by about 58%.
She expects that to keep falling.
What has not happened is Google dying. It remains one of the largest organic channels for her clients, and ChatGPT is a second destination rather than a replacement.
What changes is the metric. Clicks were the old scoreboard, and clicks are the thing being removed.
The metric now is visibility. Someone can see you in an AI Overview, again on a third-party site, again in a ChatGPT recommendation, never click once, and arrive at your signup page already decided.
How would you know today if that were happening to you?
The goals split by surface. On Google: rank as high as you can, get pulled into AI Overviews, get surfaced in AI Mode. On LLMs: get mentioned, and get cited.
The important part is that this is one content program, not two. You are building a single piece of content that has to work on both surfaces.
The Scale Trap
The most common instinct at early stage is to substitute volume for headcount. We don’t have people, so we’ll scale with AI.
Google’s spam update at the end of August is what that instinct ran into. Blogs producing unresearched AI content at volume, built to game the system rather than teach anyone, lost significant visibility. Some will take a long time to recover.
What you can take from it is that this is a time investment, not a button you push.
Formatting Is the Floor, Not the Strategy
Ask any model how to make content GEO-friendly and you get the same list. Add a summary. Add key takeaways. Use H2s and H3s. Short sentences, short sections, an FAQ block.
All of it is fine. All of it is necessary.
None of it is why you’re invisible.
Her point is that if visibility were one reformatting prompt away, every company would already have it.
Two things actually decide whether you get surfaced.
Information gain. You cannot rewrite something the web has already said ten times. If there is nothing to differentiate you, and whoever published first has more authority, there is no reason for Google or an LLM to show you instead.
Content effort. Not hours spent. Visible evidence inside the page that unique value went into it: original data, real examples, something that could only have come from you. Irina noted that Google is reportedly building an internal LLM to evaluate exactly this, and that such a system can tell the difference between a generated article and one somebody actually built.
Where the Right Topics Actually Come From
The old-school move is to open a keyword tool and start there. The flaw is that a keyword tool shows you volume, not whether your audience is the one searching.
Start instead with your sales conversations and demo calls.
Export the transcripts and run the analysis in Claude, or connect the call tool directly through MCP. Irina does this with client accounts, asking for a hundred topics and problems drawn from the clusters already in mind.
That is raw material your competitors do not have.
No sales motion yet? The data still exists. Support tickets and live chat. Onboarding and churn survey free-text. Community threads, Reddit and Slack groups. Search Console queries you rank for by accident. Review sites, where she pays particular attention to the one and two star reviews on competitors, because an angry customer naming the missing feature is about as close to a source of truth as you will find.
Three things to extract. The problems customers are trying to solve. The pain points, in the words they actually use. And the competitors they name, which is frequently not the competitor you name.
She sees that gap often. A founder identifies a main competitor and wants the keyword strategy reverse-engineered against it, and then the call transcripts show customers never mentioning that company at all. Claude connected to Gong (or your favorite AI meeting notetaker) across hundreds of calls can produce a ranked list of competitors by mention volume, which can be the basis of content for the "alternatives" articles.
Then cluster, rather than list. One article will not make you visible. Group the extracted problems, and at the beginning pick three clusters, five at most, not twenty or fifty. Authority comes from covering one problem from many angles, and you finish a cluster before expanding into the next. For a finance automation client that meant clusters like finance engineering, scaling, and AI accounting.
Only then do the keywords come in. Turn the prompts into search queries, validate the volume, and for an early-stage company set a difficulty ceiling, because a new sales automation tool is not going to outrank HubSpot.
The order is the thing worth stealing. Customer calls, then problems and pain points and competitors, then clusters, then volume validation. Keyword research lands near the end of the process rather than the start of it.
The Brief Is 80% of the Work
A good topic and a bad brief still produces an invisible page.
Alongside information gain, a brief needs consensus: the topics, entities and questions that every currently ranking page covers. Look at the top three results and find what they repeat. Check what AI Overviews cite and how those answers are structured. Miss the consensus and you don’t get considered a credible answer at all.
Consensus gets you into the running. Information gain is what you add on top. Publish only consensus and you have given nobody a reason to pick you.
Her briefs run six to ten pages and split into two halves.
The research half: the prompt and keyword set, the current answer, the consensus checklist, and the information gain requirement.
The execution half: the exact structure down to H2s and H3s, the product-led angle, internal and external links, brand and competitor framing, and supporting assets. Writers are good at making words engaging, so the structural work belongs to whoever owns search, not to the writer.
Her agency Seen automated the consensus research with an internal tool, but Irina recommends doing it by hand first, because the manual version is how you learn what the automation is doing.
The product-led angle means both positioning and actual product presence. On an account selling global tax compliance software, the goal is that ChatGPT answers “what is the best global tax compliance software” with their client, so the brief specifies the wording the writer should use to establish that association.
She calls this closer to the SEO and GEO job than the writing job. The counterweight, which came up when we discussed how naturally founders drift into talking about their own product, is that the content still has to be worth reading. You need to provide value rather than be promotional.
Publishing Isn’t the Last Step
When a model doesn’t know much about your company, it does not stay quiet. It reaches for the nearest well-documented thing: a competitor, the category average, or an old version of your product.
Irina’s habit is to reverse-engineer it. When a competitor shows up as the number one recommendation in ChatGPT, click the citation. Often it leads to that competitor’s own blog, on an article where they called themselves the number one tool.
The narrative also does not live only on your blog. Your blog carries topical authority, your help center explains how the product actually works, and your product pages say what it is and who it is for. All three are training material.
When Nextiva launched a product called Expert, ChatGPT had nothing to work from and produced answers that were simply wrong. The fix was an in-depth product page, a detailed help center, and blog articles. Within a few weeks the model was returning accurate information.
This is also the argument for the content founders tend to avoid. Best-of roundups, alternatives pages, versus pages, reviews. If you don’t publish them, competitors will, and their version is what gets picked up.
The caveat is that models now assess bias. An article that says you are the best and your competitors are irrelevant is unlikely to be cited at all. What works is a listicle grounded in evidence that didn’t come from you, such as what customers actually said on G2, with the screenshots and specifics included. One client’s listicle on their own blog now appears in AI Overviews with the citation pointing back to them rather than to a competitor.
Her rule for whether an article is worth writing at all: if you cannot naturally include your product in it, don’t write it.
The Uncomfortable Truth
Competing with established companies is genuinely hard, and studies consistently show popular brands get mentioned more often. You will need to do more than they do.
Two things keep working after the page is published.
Mentions. Third-party brand mentions are essential for AI search. For traditional Google SEO you can push good content along with backlinks and it works. For LLMs, Irina puts it at roughly fifty-fifty between your own content and what exists about you elsewhere. Get featured in listicles, and get mentioned in relevant articles about your customers’ problem, where even a short mention performs well. Asked directly whether backlinks or brand mentions matter more for LLM visibility, she didn’t hesitate: brand mentions. Backlinks still matter on Google in competitive categories, but visibility in AI search without third-party mentions is close to unreachable. This is where link building and organic growth on Reddit stop being separate disciplines and start being the same one.
Freshness. Cited content is content that stays current. Seen updates client content about twice a year, and immediately when something is underperforming. Changing the publish date is not updating.
What to Expect, and How to Measure It
This brings back the third problem, the one about expectations.
It is slow, and it compounds. In the first two months you will likely see nothing. Then first rankings, first citations, and from there it accelerates. The variable that decides the outcome is consistency.
Do not publish hundreds of pieces a month. Get the quality right, then scale.
Measurement is complicated by multi-search behavior. The same person asks ChatGPT something informational, then goes to Google, then somewhere else, which leaves attribution broken.
The simplest and most accurate method she offers is also the least technical: when someone books a demo or signs up, ask them where they heard about you.
Beyond that, watch your citation rate in a GEO tool. Irina uses Scrunch and Profound herself, though she notes that for basic tracking most tools will do, and the choice comes down to price and how much functionality you actually need. When LLMs start citing your own content, that is the leading indicator, and it arrives before revenue does. Some of this overlaps with what we covered on answer engine optimization and in our startup guide to AI visibility.
Writing for Humans, and Writing When You Are Busy
Asked whether content should be written for AI or for humans, her answer was immediate. Humans, always. The point of the content is to teach someone something and show how your product solves their problem.
Content produced purely for AI visibility does not survive contact with the next algorithm update, and she does not think readers enjoy it either. A visitor who likes your product page and then finds your articles hollow has learned something you did not want to teach them.
For founders who are busy or who don’t consider themselves writers, her advice is to change the planning horizon. Instead of researching weekly or monthly, spend one day planning a quarter of keywords and prompts, then stop returning to it. Execution gets easier when the plan already exists, and consistent slow coverage of one cluster beats one article now and another in three months.
Two things worth adding from our side. Most founders are already the expert, so the problem is extraction rather than knowledge, and speaking is usually faster than writing. A transcription tool and a few minutes of talking gets further than a blank document. You can also hand your brief to Claude and ask it to interview you, then shape the answers into the draft, which is a close cousin of the founder-led content machine approach we covered last month.
The takeaway Irina closed on was a single line. Visible content needs information gain, information gain comes from something only you have, and what only you have is your customer conversations.
