A client came to Anastasiia Krasiukova with a brand-new product, a brand-new site, and no SEO history at all. Her team wrote a handful of articles. Within the first month those articles were producing leads — from ChatGPT, not from Google, which at that point had returned nothing because the pages had only just gone up.
That gap is the interesting part. The conventional answer to “when will SEO work” is three months at the earliest. Citations in AI assistants can arrive before rankings do, which changes what a first quarter of content work can reasonably promise.
What “high quality” means now that AI can do the research
I pushed her on this, because “high-quality content” has become a phrase that survives any amount of inspection. AI has largely absorbed the job the traditional content writer did — read ten sources, synthesise, produce something more complete than any one of them. If that’s the part that’s automated, what’s left?
Her answer was specific, and all of it is work a model can’t do on its own:
- Research done by a person, with the article built on it rather than assembled from the top ten results.
- Comments from the product owner or a technical specialist at the client, folded into the piece. This is the part that makes it original rather than a better synthesis.
- Original graphics — infographics built from their own content, so the asset doesn’t already exist elsewhere on the internet.
- What people are actually saying on Reddit about the topic, quoted in the article.
- The answer in the first paragraph, and structure aimed at the fan-out queries an assistant generates rather than at a single keyword.
Notice how much of that is sourcing rather than writing. The differentiator isn’t prose quality — it’s access to a person who knows the product.
The AI experiment that actively broke things
The more useful half of the conversation was what didn’t work.
Her team had a reasonable hypothesis: use a model to re-optimise existing articles by having it flag errors, then have a content specialist work through the list and correct them. Plenty of teams are doing some version of this.
The problem was that the errors weren’t real. The model invented them, and acting on the list introduced wrong information into articles that had been correct. They’d made the content worse by trusting a review they hadn’t verified.
“We shouldn’t trust AI too much.”
She also pointed at a narrower failure mode worth knowing: on a technically complicated SaaS product, models reliably add features and behaviours that don’t exist. Every time they tried it, something fabricated appeared.
What works better: ground the model. When her team starts with a client they upload documents about the product and the niche, and instruct it to draw only on those. It still wanders occasionally, but far less.
Worth sitting with the economics, too. One of their specialists takes two to three days on an article. Generating one takes ten minutes. That ratio is the whole argument, and it only pays off if the two days buy something the ten minutes can’t — which brings you back to the expert comments and the original data.
What assistants actually cite in SaaS
Her team sets up prompt tracking at the start of every project, then looks at which sources get cited for that specific niche. The pattern:
“Listicles get cited everywhere.”
And for SaaS specifically: Reddit, G2, Clutch, Medium, plus listicles again.
The uncomfortable implication is that most of the surface area is not your website. If assistants are quoting third-party roundups and review platforms when someone asks about your category, then being present on those is the work — which is the same conclusion Saket Aggarwal reached from the link building side.
How she builds prompt tracking
Concretely, since “do prompt tracking” is advice that usually arrives without a method:
- Tooling: SE Ranking, because the volume of prompts makes manual checking impractical.
- Prompt sources: Google Search Console queries of more than eight words — long questions real people typed are the closest thing to a prompt you already have data for. Plus topics pulled from Reddit and Quora.
- Then the fan-out queries those prompts generate, tracked alongside the originals.
That GSC-queries-over-eight-words trick is the most immediately usable thing in the episode. You already have the data.
Proving it when branded and organic blur
This is the part every in-house SEO recognises. Her honest position is that separating branded from organic is genuinely hard and partly unsolvable.
What her team does instead is work backwards from the client’s own inbox: the client forwards the month’s requests, and they match those against pages the agency built. A B2B client kept receiving enquiries about an Uber API — and they had published an Uber API integration page a few months earlier. That’s not statistical attribution, but it’s evidence.
I mentioned the approach that finally worked for me: self-reported attribution (“how did you hear about us?”) reconciled against the CRM. In HubSpot I had roughly 40% of leads sitting in “direct”, which tells you nothing. Comparing the self-reported answers against that bucket left only about 4% genuinely unattributable, with most of the rest resolving to Google or AI assistants — and paid removable because I had the ad data separately.
Her counterpoint is that founders usually know anyway. As a founder herself, she can tell which leads came from AI search because they send her screenshots of the ChatGPT conversation that found her.
What she’d unlearn after ten years
Ten years ago her team tracked keywords and reported keywords. Keywords were the KPI.
She’d drop that now, for two reasons. Zero-click results mean a ranking no longer implies a visit. And the thing clients want has not changed: leads, then purchases. Keywords and traffic are leading indicators of those, not substitutes for them.
The second thing she’d change is where the work happens. Starting the agency again, she’d put more weight on external platforms — Reddit, G2, Clutch, Capterra, listicles — sometimes more than on the client’s own site, because that’s where the citations come from. Some of her clients now see roughly half their traffic from Google and half from AI assistants.
FAQ
How quickly can content get cited by ChatGPT? Faster than it ranks. A new product with no SEO history saw leads from ChatGPT inside a month, while the same pages had produced nothing in Google yet.
What does ChatGPT cite most for SaaS? Listicles in almost every niche, plus Reddit, G2, Clutch and Medium.
Can you use AI to fix mistakes in existing articles? Her team tried and stopped — the errors it flagged were mostly invented, and acting on them introduced real errors.
How do you measure results from AI search? Prompt tracking for the trend, the AI-assistant traffic segment in GA4, and leads as the KPI.
The takeaway
If you want to be cited, two things matter more than the writing: a person at the company who can say something only they know, and presence on the third-party pages assistants already quote. Content that’s a better synthesis of the existing top ten has no advantage over a model that can produce the same thing in ten minutes.
And if you’re being asked to prove any of it, start with self-reported attribution. It’s cheap, and it converts the “direct” bucket into something you can actually defend.
That’s the kind of thing I’d look at first in an organic program. If you want a straight answer about yours, book a call — and for the full conversation with Anastasiia, watch the episode or connect with her on LinkedIn.