Why OpenAI is forced to enter the search business

Or, how Google tricked OpenAI into doing it

In the midst of the buzz over the GPT-4o release and Google I/O, what’s not talked about enough is OpenAI’s intentions about entering the search market, and competing directly with Google.

This shift from “pursuit of AGI” to “commercialization of AI” surely occurred last year, and it may have contributed to the recent departures of key executives like Ilya Sutskever. It will also shape the future of GenAI - which will most likely be pervasive, multi-modal, and ad-supported.

So in this post, I’ll discuss why OpenAI is entering search, and argue that this decision was basically a forced hand ****due to Google and Meta sandwiching OpenAI from both consumer and enterprise fronts:

• Google is planning to flood all of its products - including consumer facing search - with LLM powered features. This effectively commoditizes the “ChatGPT-experience”, makes it pervasive, and steals the “AI mindshare” away from OpenAI. This puts OpenAI in a tough spot because it has far smaller user base than Google.

• Meta’s commitment to fostering the OSS Llama ecosystem, which significantly deters OpenAI’s enterprise traction by providing a private alternative.

• Most importantly, both Google and Meta demonstrated that they can produce GPT-4 level LLMs and give it away for free, which reduces OpenAI’s long term pricing power (hence OpenAI announced it’s giving away GPT-4o for free).

• Thus, OpenAI basically needs to bet on gaining decent marketshare on search to justify its valuation and expectations - among other things.
  
  

Thus, I argue that the GenAI game has changed completely to free and ad-supported, forcing OpenAI to compete in search just to stay relevant, since there’s clearly limits to how much synthetic data + one-off data partnerships can help with LLM training. If OpenAI fails to win the consumer market, then it will become “token poor” versus Google and Meta in the long run, and eventually lead to further losing bargaining power with Microsoft.

I also offer thoughts on the following questions:

• Can OpenAI beat Google in Google’s home turf, search?

• Will Apple come to OpenAI’s rescue?

• Will OpenAI releasing GPT-5 change the dynamics again?

• How does GPT-4o fit into OpenAI’s new consumer-first strategy?

Why OpenAI is entering the search market

OpenAI entering search seems like a financial no-brainer, when you consider Google’s market cap (+$2 trillion) versus OpenAI’s valuation ($80bn). Winning search for OpenAI can seem like a multi-trillion dollar call option. But there are more important non-financial reasons.

OpenAI already tried the platform strategy with ChatGPT plugins and GPT Store already, but both initiatives have failed to be commercially viable. ChatGPT Pro subscription growth has been slowing for a while now. Thus, the new growth had come to come from somewhere, and going for the search market makes sense given that sizable fraction % of ChatGPT usage is basically search anyways.

But OAI’s decision to enter search isn’t driven by profit motive, but is rather a defensive reaction, as Google and Meta have recently posed OAI significant challenges in both enterprise and consumer markets.

Difficult enterprise market: OAI is losing mindshare among enterprise customers. As Google and Meta (quite brilliantly) demonstrated in the past 6 months, OAI won’t be the only ones monopolizing GPT-4 grade LLM. This assurance of non-monopoly is huge, because:

• At least in Q4 2023, there were significant concerns as to whether OAI possessed any algorithmic secret sauce to achieve GPT-4, and it turns out others (Google, Anthropic, etc) can also catch up using the transformer paradigm.

• People also doubted whether Google can “get its act together” to ship products, and it turns out OAI re-energized Google’s productivity.

• This further encourages enterprise customers to consider multiple proprietary models, which puts OAI at a significant disadvantage. Even right now, Azure is capturing most of enterprise revenue from OAI’s tech, and this will get worse.

• In fact, other model providers (Google, Anthropic, Meta) even have some points of differentiation versus the GPT models in dimensions like context length. This suggests that OAI’s actual lead shrunk from ~2 years to somewhere around ~6 months.

• There’s already significant ChatGPT penetration among “knowledge workers” leading to tapering growth in user base and ChatGPT enterprise customers.

I also argue that OAI releasing GPT-5 may not turn the tables around immediately, though it depends on the magnitude of performance gain from GPT4 → GPT5, as well as how quickly Google, etc, catches up.

• If GPT5 beats GPT4 by how much GPT4 beat GPT3.5, then the momentum might shift again, but the market will also expect Google to reproduce it within 6 months.

• But more importantly, GPT-5 may not matter right away because companies are struggling to even adopt GPT-4 for production, and for the reasons unrelated to LLM performance (e.g. governance, internal politics, lack of urgency, etc).

Difficulties in the consumer market: But more importantly, OAI is facing much stiffer competition in consumer market, due to Google and Meta aggressively shoving as much GenAI features into its products for free. This creates a risk for OAI that consumers might get addicted to some free GenAI feature inside Gmail or Google Docs (that people are already using) and bypass ChatGPT altogether, creating top-of-the-funnel challenges.

• Google is preemptively flooding the SERP (search engine results page) with GenAI summaries, even at the risk of advertiser pushback, which indicates Google’s taking Perplexity and ChatGPT very seriously. Ideally, you want the market leader to stay distracted or complacent, but that’s not the case here.

• Google isn’t stopping there, but turning google.com into a chat product (from a search product). This means google.com is now the most popular GenAI-native app, not ChatGPT - just by definition.
  
  

Again, Google and Meta want to maintain a tight grip over the consumer “top-of-the funnel” for all things GenAI. It doesn’t even matter whether consumers actually “want” these GenAI features or not. The whole point is to take differentiation away from ChatGPT (and Perplexity to a certain extent).

The bottleneck’s data, not compute: If Google and Meta succeeds in keeping consumers away from OAI, OAI will be in a tougher position to maintain long term advantage in LLM research. Thus, OAI needs to have expand its distribution asap to avoid this fate, and it explains why OAI is launching desktop app and giving GPT-4o features away for free.

While inference and training costs are dropping 50% every quarter due to algorithmic breakthroughs and optimizations, the only constant is that there’s only so much “good, rare data” to train on. Synthetic data only helps to a degree, due to distribution and domain drifts. Google - with Youtube - has a decisive advantage here, although it’s unclear whether OAI or Meta are also using Youtube to train models.

Ironically, OAI’s upcoming partnership with Apple may all it takes to keep OAI relevant in consumer GenAI. For example, a significant win for OAI will be some exclusive partnership with Apple to offer GenAI features (like GPT powered Siri). But given Apple’s commitment to privacy, it’s doubtful that OAI will get to keep Apple device user’s data for training purposes, so I’d speculate that the partnership won’t be as deep as advertised.

Can OAI beat Google in search?

OAI entering search is - in some sense - playing right into Google’s hands. Google already has all the data partnerships with publishers, as well as the most valuable Internet asset (Youtube). For OAI to compete with Google, it has to effectively reverse engineer much of that distribution - either by themselves, through Bing, or Apple.

When it comes to search personalization, OAI cannot beat Google, which already knows way too much about everyone on a 1st party basis. Google also has local search and maps, which are probably my personal #1 and #2 use cases for using Google these days - and this dataset is not something any company can replicate easily.

In other words, for OAI to achieve feature parity with Google, OAI needs to build distribution and relationships with users, which is no longer a hard technical problem but a business problem. That’s not where OAI’s strengths lie. And I’m not sure if OAI can build distribution quicker than Google can build its own GPT-5. Besides, this work of building distribution is not something that gels with OAI’s highly technical, startup culture.

But if OAI merely copies Google’s GenAI strategy, then that’s also bad, because then it makes OAI boring. If OAI creates a similar search experience as Google - but only 10% better - then I’m doubtful that OAI can make a meaningful dent in the search market outside of the SV / tech industry audience.

Here are some other challenges and questions confronting OAI:

• OpenAI tried search already (kind of) with web search plugin - but that didn’t gain much usage among ChatGPT users. So it’s doubtful whether web search is what moves the needle for OAI.

• Bing also rolled out GenAI search - in fact, Bing was the first adopter of GenAI search - but that didn’t lead to meaningful growth for Bing.

• It’s unclear whether people will click on links on LLM generated search pages - and if that’s the case, then how will OAI get paid by advertisers?

What OAI Should Do - Keep Things Simple

OAI’s main focus should be avoiding falling into Google’s trap and try to replicate Google. OAI should instead induce Google to make mistakes in rolling out GenAI features, and capitalize on Google’s mistakes by attracting advertisers and publishers who are disgruntled with Google.

Google’s Gen AI search runs the risk of introducing too much clutter into SERP (search results page), which may turn off people who are already suffering from information overload and would prefer less information.

But if OAI just went Google’s path and messed around too much with generative UI, information cards, embedded widgets, etc - then OAI’s search product will suffer from the lack of differentiation.

People are overloaded with information already, and crave simplicity. That’s probably the reason why Perplexity, which is essentially a OAI wrapper, resonated with techies - it was slightly more helpful than ChatGPT for research purposes, but didn’t overcomplicate the UI with random info that no one cares about.

Lastly, partnership with Apple or Samsung could be a lifesaver for OpenAI - imagine ChatGPT acting as some embedded device LLM agent. That would kill numerous AI startups, but will be great for OAI as it makes ChatGPT indispensable.

Long story short, OAI should keep things simple and avoid executing blindly on Google’s GenAI strategy - simply because I don’t think OAI even has the resources to fight so many wars on all fronts that Google competes in. Heck, OAI should even steal Perplexity’s playbook. Otherwise, if OAI tries to become Google, then that’s exactly playing right into Sundar’s hands and ChatGPT will lose a lot of its luster.