{"id":4976,"date":"2026-08-09T01:45:04","date_gmt":"2026-08-08T22:45:04","guid":{"rendered":"https:\/\/ivitskiy.com\/blog\/definitive-ai-seminar-2024\/"},"modified":"2026-08-09T01:45:04","modified_gmt":"2026-08-08T22:45:04","slug":"definitive-ai-seminar-2024","status":"publish","type":"post","link":"https:\/\/ivitskiy.com\/blog\/en\/definitive-ai-seminar-2024\/","title":{"rendered":"Overlooked Google Ads Insights: Igor Ivitskiy at Definitive AI Seminar 2024"},"content":{"rendered":"\n
In short:<\/strong> A Google Ads account should become efficient before it becomes larger. Drawing on insights from 700 million in client ad spend, Igor Ivitskiy organised the work into three phases: cut waste, tune the system, then scale. One company pushed volume to 300,000 leads a day, but 90% were fake; in another account, budget reallocation lifted profit by 30% without a budget increase. The talk argued that business outcomes, not interface prompts, must verify every decision.<\/div>\n\n\n\n

Definitive AI Seminar<\/a> took place in Chicago, United States, on April 16\u201318, 2024, as a live conference seminar. In a stage session, Igor Ivitskiy<\/a> presented lessons from his work with Google Ads.<\/p>\n\n\n\n

\u201cOverlooked Google Ads Insights\u201d<\/strong> connected search-term and keyword research<\/a>, placement quality, product economics and audience data within one strategic sequence. The subject was not a single campaign feature. It was how to decide what deserves more budget and what should be questioned first.<\/p>\n\n\n\n

+30%<\/div>
profit in one client account after reallocating budget geographically, with no increase in total budget<\/div><\/div>\n
Source: Definitive AI Seminar 2024 talk, Igor Ivitskiy.<\/div>\n\n\n\n

Scale begins with the quality of the system<\/h2>\n\n\n\n
\"Dashboard
Dashboard where two overlooked tiles are highlighted<\/figcaption><\/figure>\n\n\n\n

Scaling multiplies whatever an account already contains. If the conversion signal mixes paying customers with fake leads, more budget buys more of both. If weak placements or products consume the budget, a larger campaign gives them more room. Cutting waste comes first because it changes the quality of the evidence used in every later decision.<\/p>\n\n\n\n

This is why apparent relevance is not enough when evaluating search traffic or negative keywords<\/a>. A word that looks commercially weak may still appear in profitable searches. The same caution applies to platform guidance. Google describes optimisation score<\/a> as an estimate of how well an account is set to perform, while the talk treated every recommendation as a hypothesis that still needs a profit test.<\/p>\n\n\n\n

Optimisation then becomes a recurring management process rather than a cleanup project. Google presents Ad Strength<\/a> as feedback on the relevance, quality and diversity of ad assets. Igor\u2019s account data showed why such feedback should not be confused with a business result. Even tools such as broad match<\/a> can support growth, but only after the account has a reliable way to judge lead quality, cost and profit.<\/p>\n\n\n\n

\n
1<\/span>Cut waste<\/strong>
Challenge search traffic, placements and product spend with evidence from the business.<\/div>\n
2<\/span>Tune the system<\/strong>
Turn geographic, device and audience differences into a continuous optimisation process.<\/div>\n
3<\/span>Scale deliberately<\/strong>
Expand only after the account can distinguish useful growth from extra noise.<\/div>\n<\/div>\n
Source: three-phase framework presented at Definitive AI Seminar 2024.<\/div>\n\n\n\n

What I want to know before scaling<\/strong><\/p>

I do not treat more clicks or leads as proof of growth. First I want to know which traffic produces real customers, where the budget is being wasted and whether the account can keep learning from those differences.<\/p><\/section>\n\n\n\n

This decision discipline is easier to understand when it is applied to complete account data. The school\u2019s training uses the same principle to connect campaign metrics with the economics behind them, without treating any interface recommendation as a promise.<\/p>\n\n\n\n\n\n\n\n

Key insights<\/h2>\n\n\n\n
    \n
  • Scale amplifies signal quality.<\/strong> A company reached 300,000 leads per day after pushing for rapid growth, but 90% of those leads were fake.<\/li>\n
  • Obvious labels can mislead.<\/strong> People searching with the word \u201cfree\u201d often bought, so Igor rarely rejected that word without analysing its actual economics.<\/li>\n
  • Product portfolios hide budget competition.<\/strong> In the e-commerce pattern described on stage, roughly 80% of products were \u201czombies,\u201d 3\u20135% were strong performers and about 15% still needed evidence.<\/li>\n
  • Platform scores are hypotheses.<\/strong> In a comparison with close to 600,000 observations per ad, the ad rated \u201cpoor\u201d produced twice the CTR and 1.5 times the conversion rate.<\/li>\n
  • Reallocation can outperform expansion.<\/strong> One account increased profit by 30% without raising total budget after spend moved towards stronger geographic areas.<\/li>\n
  • Optimisation is a process.<\/strong> Markets, competitors and account data change too quickly for a one-off audit to remain sufficient.<\/li>\n<\/ul>\n\n\n\n

    Key moments from the talk<\/h2>\n\n\n\n

    Scepticism as a working method<\/h3>\n\n\n\n

    Igor opened with a question learned from MythBusters: \u201cIs it true?\u201d He used it well beyond advertising. His family had believed for a century that his grandfather had disappeared in action after being marked an enemy of the Soviet regime. A search across spelling variants led Igor to a New York arrival record, census traces and eventually a grave near the city. The story established the talk\u2019s standard of evidence. A familiar explanation may feel settled because it has been repeated for years, but repetition is not verification. In a Google Ads account, the same scepticism applies to inherited limits, default reports and apparently self-evident assumptions about user intent.<\/p>\n\n\n\n

    The advertising evidence behind the session came from 700 million in client spend accumulated across Igor\u2019s work. Yet the purpose of that scale was not to present a universal formula. It was to show how often a plausible label fails when tested against behaviour and profit. The word \u201cfree\u201d was his simplest example. Advertisers often assume that anyone using it cannot become a buyer, but his data repeatedly showed purchases from those searches. Search terms therefore have to be judged by their contribution to the business, not by the emotional reaction a word creates. When profit data is still thin, engagement and competitive signals can offer an interim view, but Igor clearly treated them as substitutes for missing evidence, not equal proof.<\/p>\n\n\n\n

    What rapid growth exposed<\/h3>\n\n\n\n

    The warning against premature scale came from a company generating about 30,000 leads per day. Its team wanted one million and asked mainly for growth ideas. Igor first urged them to understand which leads became paying customers and which were fake, but the analysis did not happen. Three months later, the account was producing around 300,000 daily leads and 90% were fake. Their summary was blunt: \u201cWe just scaled.\u201d The campaign had achieved more volume while weakening the meaning of that volume. The recovery began by looking for behavioural differences between real users and fake leads, then moving existing budget towards the former. The strategic lesson was larger than this case: a target metric becomes dangerous when the system can satisfy it without delivering the business outcome behind it.<\/p>\n\n\n\n

    Waste also appeared in less visible forms. Some display sites, video channels and apps existed mainly to collect advertising revenue and contributed little to the advertiser\u2019s result. In e-commerce accounts, large catalogues created internal competition for budget. Igor described a recurring distribution in which about 80% of products behaved like \u201czombies,\u201d consuming spend while performing poorly; only 3\u20135% were established strong performers, and roughly 15% remained undecided because they needed more evidence. That does not make a fixed product rule. It shows why an aggregate campaign total can conceal opposing economic roles. More budget cannot correct that structure by itself. It may simply let weak inventory continue to crowd out products whose demand has already been demonstrated.<\/p>\n\n\n\n

    From isolated changes to a learning loop<\/h3>\n\n\n\n

    The next challenge was separating platform guidance from commercial evidence. Igor compared two responsive search ads with close to 600,000 observations for each. The ad labelled \u201cpoor\u201d because it used fewer assets produced twice the click-through rate and 1.5 times the conversion rate of the comparison ad. The result did not prove that low Ad Strength is desirable. It showed that a diagnostic score and an advertiser\u2019s profit answer different questions. The same distinction applies to optimisation recommendations and sales calls encouraging automation or higher budgets. They can generate testable ideas, but the account owner remains responsible for deciding whether the ideas improve qualified demand, conversion quality and profit.<\/p>\n\n\n\n

    Igor\u2019s alternative was continuous optimisation. In one account, reallocating spend among geographic areas raised profit by 30% without increasing the total budget. The important point was not the particular geographic setup, which he explored in the full session, but the operating rhythm behind it. User behaviour, competitors and available inventory keep changing, so an adjustment that works now will not protect an account indefinitely. Only after waste has been reduced and the learning loop is functioning does deliberate scale make sense through wider reach, new campaign formats or other traffic sources. As Igor put it, \u201cIt\u2019s not a panacea. It\u2019s a process.\u201d Google Ads can be a major source of valuable users, but its value depends on repeated measurement rather than a single optimisation event.<\/p>\n\n\n\n

    The full recording of this talk is included in the bonus pack for students.<\/p>\n\n\n\n