The origin story of Cognitiv contains a pleasingly efficient piece of product feedback. Jeremy Fain and Marc Hudacsko wanted to leave their jobs and build a mobile advertising platform. They recruited their childhood friend Aaron Andalman, a neuroscientist who had studied learning in zebra finches at MIT and neural circuits in zebrafish at Stanford. His review, as the founders later recalled on a podcast, was blunt: the mobile-DSP idea was dumb. He wanted to do deep learning.
Fain and Hudacsko had barely heard the term. They listened anyway. In 2015, before every earnings call acquired an AI paragraph, the three fifth-grade friends founded Cognitiv around a harder and less fashionable proposition: neural networks could make better advertising decisions than the standard bidding algorithms Fain had watched circulate through the industry.
The complaint was specific. Target and Walmart do not have the same customers, yet ad platforms kept handing different brands similar optimization machinery. Even custom algorithms demanded a team to pick variables and maintain rules. Cognitiv's answer was to train a neural network for the advertiser's actual outcome, let it absorb more signals than a conventional model, and update the bidding logic as behavior changed. The thing it sells is not an AI-written ad. It is the decision, made before an auction closes, that this particular impression is worth buying.
A model in every aisle
Cognitiv now sits in a useful middle layer of the advertising market. It is a demand-side platform when a customer wants a managed service. It is a modeling and curation partner when an agency wants to keep its hands on The Trade Desk, Google Display & Video 360 or another DSP. And it is a set of specialist products when the problem is contextual relevance, connected television or audience creation.
| Product | The job | How it ships |
|---|---|---|
| Deep Learning DSP | Optimize impressions toward a business KPI | Managed media buying |
| Curation | Select likely converters before a bid reaches the buyer | Dynamic private-marketplace deals |
| ContextGPT | Find pages whose meaning and sentiment fit a brief | Chat planning, then any DSP |
| Performance CTV | Target TV viewers and measure online actions | Cross-device campaigns |
| AudienceGPT | Build an audience from current journeys, not an old label | DSP, Deal ID or data segment |
That flexibility is part of the business model. Cognitiv runs enterprise campaigns and sells managed services, but it also packages model output into Deal IDs that travel through existing buying systems. Public list prices are not part of the pitch; this is a sales-led product tied to media, integrations and campaign objectives. The company's disclosed funding is modest by contemporary AI standards - about $3.25 million in the supplied company record, including a $450,000 convertible note in 2019. It has had to make distribution do real work.
From sentence to spend
The shirt that was never for sale
AudienceGPT explains the company's thesis with a stained shirt. Someone visits a clothing brand's product page to check fabric and washing instructions. A retargeting system sees “shirt page” and spends the next week chasing a presumed shopper. A person sees the real story: the visitor already owns the shirt and wants to clean it. The click is a signal. It is not the meaning.
Cognitiv says its new audience product combines the surrounding context, general world knowledge and reasoning to estimate that meaning. A strategist describes a target in plain English; the system proposes a persona, estimates reach, lets the buyer adjust the balance between relevance and scale, then activates across web, connected TV, audio or social. Instead of freezing someone in a taxonomy for months, it can add and remove people as their apparent journey changes. The company says audiences can refresh within 15 minutes and that it evaluates browsing behavior across more than 250 million people.
An audience is no longer a static list. It is a living hypothesis, one that evolves as consumers do.Aaron Andalman, co-founder and chief science officer
The phrase “living hypothesis” is more honest than the usual targeting language. No model can read a shopper's mind. It can make a better, fresher probabilistic guess, show its reasoning and measure what happened. For marketers, that means a ski-equipment brief can expand beyond people who once entered a “winter sports” segment and toward people currently reading about snow gear, mountain travel and outdoor plans. For a movie studio, a Cognitiv case study describes placing ads beside discussion of a competing release, finding moviegoers at the moment opening-weekend choices were being made.
Read the label: These are Cognitiv's company-reported 2025 operating figures, not audited public-company results. Its case studies are similarly directional: Welch's reported 47% higher engagement than its creative benchmark, while an unnamed auto retailer reported 50% lower cost per lead.
The clever model was too slow
Cognitiv did not reach the chat interface in a straight line. The original ContextGPT used similarity scoring and a collection of sliders to steer abstract qualities such as sentiment. The trouble was multiplication: every new dimension wanted its own dedicated model, and simple mathematical similarity could confuse content that felt optimistic with content that merely contained the words “positive sentiment.”
A scientist named Mohsen suggested letting the prompt define relevance itself. In a head-to-head test, the method matched the dedicated slider models. That changed the product direction. A reasoning model could interpret whether the advertiser cared about audience, tone, format or semantic meaning. But the early version needed 10 to 15 minutes to process a prompt - fine for a lab demonstration, absurd for a planning tool feeding an exchange measured in milliseconds.
The team built a multilayer relevance engine and distilled the judgment of larger foundation models into faster, prompt-specific models. Cognitiv says the September 2025 upgrade improved placement accuracy by as much as 40 percent while adding conversational memory and explanations for recommendations. That same compression technique later became the technical foundation for AudienceGPT. The lesson is less glamorous than “add AI”: expensive reasoning became a product only after the company found where to spend it once and where to reuse it cheaply.
A moat made of pipes
Cognitiv competes with large DSPs, contextual specialists such as GumGum, newer AI-led buying tools and the data-science teams inside agencies. It also partners with platforms that look like alternatives. That apparent contradiction is the strategy. A brand can use Cognitiv's managed DSP, but an agency can keep its preferred controls and consume Cognitiv through a curated marketplace deal. PubMatic, Index Exchange and Magnite extend the supply path; LiveRamp distributes AudienceGPT segments; an expanded OpenAI agreement gives ContextGPT more processing capacity.
The January 2026 Magnite integration makes the infrastructure argument visible. Cognitiv's models are co-located closer to the supply, allowing more computation before the auction clock expires and feeding curated inventory into Magnite's ClearLine. Fain says this setup offers ten times the computing power of containerized alternatives. Whether that figure holds across every rival is difficult to compare, but the underlying constraint is real: a beautiful model that misses the bid deadline is decorative software.
Customers tend to be brands and agencies with large, measurable media problems: retailers hunting new buyers, pharmaceutical marketers navigating suitability, travel companies finding intent, entertainment studios fighting for an opening weekend, and CPG brands trying to connect an impression with a purchase. Named work includes Welch's, Industrious and Fanatics Sportsbook; many case studies keep the advertiser anonymous. Cognitiv employs roughly 200 people, with the company saying nearly half work in science and engineering roles.
What to steal - and when not to
The most copyable part of Cognitiv is its packaging, not its neural-network vocabulary. It began with one expensive decision - whether to buy an impression - and attached learning to a measurable outcome. It lets customers enter through workflows they already have. ContextGPT shows why it recommends a page. AudienceGPT lets a buyer widen or narrow the relevance threshold. Each product turns a research system into a control a media team can understand.
The Cognitiv playbook, in four moves
- Pick a frequent decision where a small accuracy gain compounds into money.
- Train on the customer's outcome, not an industry-average proxy.
- Ship through the customer's existing stack before asking them to replace it.
- Expose reasoning and a precision-versus-scale control so the operator stays involved.
There are conditions. Custom deep learning is a poor ornament for a tiny campaign. Fain has previously said the approach needs meaningful signal volume - around a thousand conversion events at minimum for some custom-model work, with more generally better. A vague objective produces a vague optimizer. Sparse data, short flight times and minuscule budgets may favor simpler rules or standard models. Teams that require a fully deterministic explanation for every outcome will also be uncomfortable with probabilistic machinery, even when recommendations are annotated.
Privacy-safe contextual products do not erase every privacy question, either. ContextGPT can work without cookies, pixels or personal data because it evaluates content. AudienceGPT is different: it reasons from behavioral context and distributes audiences through advertising infrastructure. Buyers still need to examine consent, identity and governance in the markets where they operate. “No seed list required” is a product capability, not a universal compliance pass.
The company remains private, with no public revenue or valuation, so the cleanest proof is adoption and repeatable customer performance. Cognitiv reported ContextGPT growth of 388 percent in 2025, 67 percent more new clients and an average return-on-ad-spend improvement of 29 percent from the first half to the second. In March 2026, Fain and Andalman landed on ADWEEK's AI Power 50; later that month, AudienceGPT launched. Awards are pleasant. The harder test is whether a marketer who described an audience on Monday buys better media on Tuesday.
That brings Cognitiv back to its original irritation. A standard algorithm treats unlike brands alike. A static segment treats a changing person as a permanent noun. The company has spent eleven years replacing both with models that keep revising the answer. It is a demanding way to buy an ad. In the cases where the signal, scale and stakes are present, it may also be the sensible one.