Writing for the machine that talks to your customers

For two decades, digital discovery meant search engines. Today, that model is changing. Consumers increasingly ask ChatGPT, Claude, Gemini and Perplexity directly for recommendations, often accepting a shortlist rather than browsing pages of links.

In this first conversation of the Making Strides series, Reha Sönmez, founder of NetRanks, explains why he believes the future of visibility is built on trust signals understood by AI models rather than traditional SEO tactics.

The changing context of search

For most of the internet's commercial history, being found meant ranking. A prospective customer typed a query into Google, scanned a page of blue links, and clicked. An entire industry grew up around it: keywords, backlinks, crawlable pages.

That way of being found is losing ground, and the change Reha describes goes deeper than the interface. "The fundamental gateway has changed," he says. "Previously we would go to Google, make a query, and read through the list. Now we just ask a question to an AI search engine, and it comes back with an answer." Where a search engine returned a menu of sources and left the choosing to the user, an AI assistant does the choosing itself. A new layer now sits between a company's content and its customer, one that ingests everything and returns a verdict. The buyer no longer browses the shelf; the machine hands them a single item.

Crucially, this is not confined to the challenger apps. Google itself increasingly answers queries with an AI Overview rather than a list, compressing what were ten links into a paragraph the user does not have to scroll past. What Reha is describing, then, is not simply a defection to ChatGPT but a change in the nature of discovery across every surface, incumbent and upstart alike.

A skeptic would reasonably ask: is this niche behaviour, confined to early adopters and a few digital-native categories? The numbers offer an ambiguous answer that Reha reads with conviction. AI still accounts for only ten to fifteen percent of total search traffic; on that measure, Google's dominance looks intact. But he points to the slope rather than the level, and the forecasts he cites for AI reaching half of all search cluster around 2027 to 2028. Adoption is uneven: he says conversations about health, finance, technology, food and travel are migrating first, with other sectors expected to follow. "If I want a doctor now," he says, "I ask Claude, not Google."

For a founder who hears "search" and still pictures a page of Google results, the implication is uncomfortable. The question is no longer where you rank. It is whether, when an AI is asked to recommend, your company is named at all.

A trust problem, not an optimization problem

The instinct of every marketer confronting this shift is to reach for the familiar. A new acronym has already appeared to reassure them: generative engine optimization, or GEO, a deliberate echo of the SEO domain it hopes to inherit. Reha thinks the word smuggles in a false premise.

The premise is that AI search can be gamed the way Google can. It cannot – at least not in the same way – because the two systems differ at the root. "Ask Google the same question twice and you get ninety-nine percent the same answer," he says. "With AI search, that is not the case; there is a non-deterministic component." A search engine is, in effect, a lookup: the same input yields the same ranked output, and an optimizer can reverse-engineer the rules and climb. A language model answers probabilistically, drawing on a shifting internal sense of what is relevant and credible. Ask it twice and it may answer differently. "Once you have that component," Reha says, "optimizing for it becomes very hard, because you have a measurement problem."

His reframe is the intellectual core of the conversation: "How an AI engine behaves is better understood as building trust: letting that model understand your industry and your brand well enough that it's more likely to recommend you. So I wouldn't call it an optimization problem. It's a long-term trust-building problem."

If there is no fixed ranking to climb, the task is not to optimize a page but to earn a standing: to make a model understand your business well enough that it reaches for you when asked. That is closer to reputation than to engineering, and it accrues slowly. The analogy Reha draws is to public relations, shaping what a third party says about you when you are not in the room.

He calls it "AI relations". The distinction changes what a company measures, who owns the work, and how long results take. Keywords and backlinks do not disappear, but they stop being the point. The point becomes whether the machine believes you, and, as Reha has found out, some of the most established brands discover it does not.

What most people still miss

Even those who accept that AI search is coming tend to underestimate it in a particular way: they model it as a swap. Google's traffic migrates to ChatGPT, the search market changes hands, and the game continues much as before with new players on the board. Reha thinks this is wrong on two counts.

The first is conversion. People use these models for far more than search – drafting, reasoning, working through decisions – so a relationship of trust is established long before any commercial question is asked. By the time a user requests a recommendation, the assistant is a familiar counsel rather than a cold directory. A referral from a source the user already believes behaves nothing like a link on a results page, and it shows up in the numbers.

"With Google, trust comes late: you read the content, you check the site yourself. With AI, because we already use these models for non-search conversations, we've established trust before we even ask. So AI-sourced users convert at a much higher rate, and that's measured."

The second error is larger, and it concerns scope.

"What people miss is that they think of AI search as just a continuation of the search market taking over. In reality, it's going to take over the entire product discovery market: advertising, third-party messaging, everything. Even when I see an ad now, I don't go to Google; I ask ChatGPT, 'is this good for me, or are there better alternatives?' So all product discovery, all information discovery, shifts to these engines. The market sizing is vastly under-calculated. And it's not only products. Any form of opinion-shaping goes there too. The next generation will ask AI which party to vote for… I'm sure some already do."

The category these engines will absorb, in other words, is not the SEO market but the whole of discovery – and beyond commerce, a growing share of how people form their views. Taken to its conclusion, the model becomes a single intermediary standing between a person and almost everything they might come to want, buy or believe: an executive assistant that screens the world on their behalf.

"The human-knowledge interaction now has an AI gateway. How it behaves is critical to anyone on the other side of the information market: how you reach your customers and your audience will be gated by AI. That's the problem we are building NetRanks to solve, the keys to that gateway. We measure how it behaves, back everything with data, and try to predict its behaviours so we can help customers navigate this new era of discovery."

For any business on the far side of that gateway, the terms of access are no longer set by a search algorithm they half-understand, but by a system whose reasoning is opaque and whose trust has to be earned. The first sign that it has not been earned is already showing up on marketing dashboards.

The canary: your SEO ROI

For all its conceptual weight, the shift announces itself in a mundane place: the marketing budget. A company does not need to understand large language models to feel the change. It only needs to watch its returns. Reha says the earliest and most reliable symptom is a widening gap between what a company spends on search and what comes back.

"Companies already spending on SEO should measure their ROI very precisely. What we hear constantly from inbound customers is, 'We spend the same money on SEO, sometimes more, but the return isn't there. We don't get as many customers as before.' Performance marketers usually get the signal first, because the traffic itself is diminishing. Fewer people scroll and click on Google. You might still rank on the first page, your SEO is fine, but people aren't looking there anymore."

The tell is that nothing appears to be broken. Rankings hold, the site still sits on the first page, the scorecard looks healthy. What has changed is upstream: fewer people are reading the page at all, because the answer they came for arrived before they reached it. This is why performance marketers register it first. They sit closest to the numbers, and theirs is the function where thinning traffic turns soonest into a thinner result. For the rest of the company the signal is deferred; for them it is already flashing. Reha's counsel is to read a flat line against rising cost not as a soft quarter but as a leading indicator of where discovery is going.

The canary in the mine did not make the air toxic; it simply stopped singing first. And for some companies, Reha has found, the diagnosis turns out to be stranger still: the problem is not that AI overlooks them, but that their own website is persuading it to recommend someone else.

Self-sabotage

The most unsettling story Reha tells is not about being absent from the answer. It is about a company that was present, trusted, and losing anyway.

When an AI assembles a recommendation, it typically reads a set of sources and often discloses which ones it leaned on. For one of NetRanks's customers, the single most-used source was the company's own website: the machine had found them, read them, and treated their site as the authority on the question. Then it recommended their competitors.

"For this customer, their own website was the top source AI used, which is great, it means they control the narrative. AI liked their site and read it. But then, after reading it, it went and recommended other brands. Their content was missing the right trust signals. So every time AI read their website, it broke its own trust and recommended a competitor. They had the power to control the answer and were using it to recommend other brands. Their own website was sabotaging their visibility."

The failure was one of persuasion, not visibility. The site had been built, like almost every corporate website, for two readers: people, who want something legible and attractive, and search engines, which want something indexable. It had never been written for an AI that reads in order to decide whom to trust. The logical conclusion is almost absurd: this company would have been more visible with no website at all. There is no agreed name for the phenomenon yet – Reha offers "negative signalling," or "trust-breaking content" – and, tellingly, it cannot be seen without the right instruments. A company cannot simply read its own pages and spot the fault, because the fault exists only in how a model interprets them. It has to be measured.

What NetRanks does

If the problem is invisible to the naked eye, the first job is to make it visible. That, stripped to essentials, is what NetRanks sells: you cannot manage what you cannot measure, and why an AI reaches for one brand and not another has until now been unmeasurable.

"NetRanks deliberately measures how AI search behaves given industries, languages and regions. We reverse-engineer how these engines behave, understand which signals they pick up on and which are actually hurting a brand's visibility, then help brands get the right signals into their content so the engines trust them more."

The method is empirical. NetRanks interrogates the engines at scale – the same question asked many times, and many variants of it, across models, languages and regions – and records what comes back: who is named, in what order, and which sources the model cites. From that it works backwards to the signals that moved the outcome. Where traditional search rewarded roughly two such signals, keywords and backlinks, Reha says his team has catalogued more than a thousand, most of them irrelevant to Google and invisible to the tools built for it. The relationship runs one way: customers who implemented the feedback saw no drop in their Google rankings. The new signal set contains the old one; it does not compete with it.

The output is unglamorous and exact: not "improve your content" but, in his own words, the signals missing from "page two, paragraph three, sentence one." Larger clients receive this as hands-on consultation; a newer self-serve product at netranks.ai lets a company gather its own data, submit its site, and get a task-by-task roadmap, with measurement running continuously in the background.

What that looks like in practice is best shown by a company that had every conventional advantage and still could not be found. It had hundreds of millions in revenue and a permanent place on Google's first page; yet across a thousand questions put to the AI engines, it surfaced two or three times. NetRanks collected the data, tied it to a single product, and returned about a hundred and thirty precise changes to one or two pages. Within days, two or three mentions per thousand had become four hundred, making the brand one of the three most frequently named in its category. Position lagged behind frequency: at first the model named it only at the foot of its answers, an afterthought, and it took a further month of iteration before the brand appeared among the first names offered. Visibility rose first, then traffic, then leads – in that order. Trust, as Reha keeps insisting, is accrued, not switched on. And once a brand learns to accrue it, a subtler question follows: what happens when the machine trusts you for the wrong thing or in the wrong place?

Positive trust, negative trust, wrong-kind-of trust

It is tempting to treat AI visibility as a single dial running from invisible to prominent. The reality Reha describes has a third axis. A model can fail to mention you; it can mention your rivals off the back of your own content; and it can name you in precisely the context you would pay to avoid. Presence is not the same as the right presence.

This last case is the wrong kind of trust: a brand surfaced for the query it would least choose, or bound to a use, a price tier or a peer group that misrepresents it. To a model doing its earnest best to be helpful, such a placement reads as success; to the company it can be worse than silence.

What is striking is that a handful of Reha's customers have already arrived at the inverse of the usual request.

"One thing we're now exploring with a few customers is: how can I make sure I'm not in the AI answers? Because I don't want to be found in certain types of answers. It requires the exact same data collection and signal detection, and we can guide a company to remove itself from certain answers the same way we get it into others."

The apparatus, he notes, is the same run backwards: the signals that earn a recommendation, once understood, can be tuned to withhold one. NetRanks claims to be alone in offering this today, and whether or not that lasts, the request itself is instructive. It marks the moment at which AI visibility stops being a growth tactic and becomes something nearer to reputation management – the ongoing work of deciding not merely whether the machine talks about you, but what it says, and where.

That work is complicated by one last fact: there is no single machine to manage, and the ones that exist do not hold still.

Different engines, different signals, and the lock-in to come

Reha is quick to puncture the hope that a single set of fixes will serve everywhere. ChatGPT, Gemini and Perplexity share some of the signals they respond to, but weight them differently, and the weighting shifts by industry. His team's approach is therefore sequential: get a client named across all the major engines first, then study each in turn and reinforce wherever the brand is weakest. The tools are engine-agnostic by design; NetRanks treats the task as signal extraction, and the same apparatus, Reha says, would work on a small language model as readily as on a frontier one. Even the tiers within a single product (the free model, the paid, the "research mode" that returns an essay rather than a sentence) can behave differently enough to change who gets named.

That fluidity is the industry's present mercy, and Reha does not expect it to last. Trust, once established, compounds: the longer a model has recommended a brand for a question, the harder a rival finds it to break in. For now, no engine has won, and every model update reshuffles the deck. Reha points to Gemini's move from one version to the next, which shortened the list of brands it would name for a given query. The effect was brutal in its simplicity: the top names gained still more visibility, and the bottom half of the answer simply disappeared. He calls it a visibility extinction.

The historical rhyme he reaches for is the early web. In Google's first years, the businesses that learned its ranking logic early paid far less for their position than those who arrived once the rules had hardened and the incumbents had entrenched. AI search, he argues, is running the same play at higher speed. "Same trend, steeper curve, and much faster." When these engines mature, he expects a lock-in more complete than Google ever managed.

The counterintuitive conclusion is that the present disorder is the opportunity, not a reason to wait. Because the ground is still moving, incumbency counts for little: a company founded this year has roughly the same chance of winning the next era of discovery as one that spent two decades mastering the last. The advantage will fall not to the largest marketing budget but to whoever begins earning the machine's trust first, and grasps that it is trust, not traffic, they are building.

The moat, and the measurement question

All of which leaves the question a persuaded reader should be asking: if AI answers are probabilistic and the measurement problem is real, what separates NetRanks from the SEO shops now rebadging themselves for the new era? Reha's answer, in effect, is that the measurement problem is the moat.

Advice differs by industry, language, region and model. What serves a healthcare brand in English says nothing about a finance brand in French, and a recommendation that fails to clear a statistical threshold is withdrawn rather than shipped. Whether the models behind that claim are as good as advertised is precisely what a customer should test. But it is the right kind of answer to the hardest question this piece raises: a system that cannot be optimized deterministically can still be measured probabilistically, provided someone does the unglamorous work of asking the same thousand questions a thousand times and counting what comes back. That work, a team that spent its career inside large-scale search and ranking systems, and the data that compounds with every measurement – that, Reha argues, is what cannot be "vibe-coded overnight".

Which returns the argument to where Reha began. The task is no longer to be found by people; it is to be understood, and vouched for, by the system that people now ask. Write for that reader, and the humans follow. Ignore it, and no amount of first-page ranking will rescue you from an answer that never says your name.

In the next conversation, Reha turns from the problem to the venture itself: how NetRanks went from an idea to a funded startup.