Why do so many LLMs suddenly become slow and flaky when you ask about things that only appeared after 2025? This post isn’t about parameters.
The reason isn’t complicated. Many models have never actually “read the full text”—they just get a title, a URL, and a snippet of a few dozen words back from a search engine, then rely on the model itself to fill in the rest. When the factual foundation is that thin, the longer the output gets, the higher the odds of hallucination.
I’ve tiered the products and search setups I’ve used based on whether they can pull the full text, whether they can trace sources, and whether that attribution is precise enough. Anyone doing research, writing analysis, or looking up new material can use this to pick their tools.
Who this is for
- People who regularly use AI to look up recent information, policy updates, and model news
- People sourcing topics, writing reports, or covering industry trends
- People who’ve already noticed that the same model can be brilliant one moment and completely off the rails the next
- People trying to decide whether to use web UIs, APIs, Research modes, or external retrieval tools
If most of what you ask about is common knowledge, history, or basic concepts, this won’t be as helpful.
Don’t conflate model capability with search capability. A lot of the time, the retrieval pipeline is just too weak; it’s not that the model itself is bad.
A simple rule of thumb:
- If it only gets snippets instead of the full text, don’t expect it to answer new topics reliably
- If it can fetch the full text but only links back to the top of the page, it works, but you’ll still have to do a lot of manual checking
- If it can grab the full text and pinpoint the exact sentence, its practical value jumps significantly
- For sites with aggressive anti-scraping like X or WeChat Official Accounts, many approaches fail by default
- For specialized content like Reddit replies or YouTube subtitles, not every search source can retrieve them
So when asking about anything new after 2025, prioritize tools that can “fetch full text + trace sources.” If it’s just a standard web search endpoint, the output might look complete, but in reality it’s likely just “100 words of information + 900 words of hallucinated filler.”
What you’ll need
- Your regular LLM portal: ChatGPT, Claude, Gemini, or Codex all work
- At least one research tool capable of fetching full text
- The habit of manual verification
- The baseline expectation that “search results do not equal facts”
Here is how I tier them right now, based on my own usage records prior to March 2026:
- Tier 1: NotebookLM Web, Perplexity API
- Tier 2: Gemini Deep Research Web, GPT Deep Research Web, Claude Research Web, Manus Wide Research Web
- Tier 3: Tavily Extract API, Jina Reader API, Exa API, Grok Web
- Tier 4: Gemini Web
- Tier 5: GPT API / Web / Pro Web, Codex, Claude API / Web / Code, Gemini Deep Think Web, OpenRouter’s :online
- Tier 6: Antigravity, Gemini API, Gemini CLI
- Tier 7: Domestic Chinese official search pipelines
I only kept track of one pricing note: the Perplexity API opened in September 2025. I haven’t calculated the exact billing breakdown, so check their official site for current rates before using it.
Follow this order
- First, determine whether what you’re asking about is “post-training new information.” If it’s a model update, product change, policy adjustment, or site rule update that happened after 2025, don’t assume the model knows it.
- Check whether your tool gets snippets or full text. If it only retrieves a dozen search results with thirty- or forty-word snippets each, treat the answer as a lead, not a conclusion.
- Prioritize options that can fetch the full text. In my experience, NotebookLM Web and the Perplexity API are closest to being truly usable; web-based Research products rank next; standard search-augmented chat products trail further behind.
- Check whether source attribution is granular enough. Jumping to the exact sentence is best; landing at the top of the article is second best—the former makes verification much cheaper. Gemini Web supports sentence-level deep links via
#:~:text=, which helps with manual verification. - Don’t blindly trust results from specialized sites. Google has an easier time retrieving YouTube subtitles and Reddit comments; for sites with anti-scraping like X or WeChat Official Accounts, many approaches simply break down.
- Finally, do your own cross-verification. Open at least 2 to 3 links to confirm that the content cited by the model is actually there—don’t just judge by how convincing the writing looks.
Common pitfalls
-
Confusing “web access” with “knowing how to research” These two things are worlds apart. Having web access doesn’t mean being able to read full text, let alone quote accurately.
-
Letting your guard down the moment you see source links Some tools only link to homepages rather than specific sentences. You have no idea how the model connected the dots in between.
-
Assuming the API is always better than the web app My experience is the exact opposite: quite a few API pipelines have weaker retrieval, and some only return temporary redirect URLs instead of real sources.
-
Using domestic Chinese search pipelines for English AI topics If a topic already has a high density of English information, the results will take a noticeable hit. That’s why I separated domestic Chinese pipelines into their own tier.
-
Mistaking model degradation for a prompt issue Sometimes it’s not that your prompting is bad—it’s that the retrieval foundation simply isn’t there. Prompts can only fine-tune; they can’t fix missing information.
My recommendations
If you’re genuinely using AI to track new developments and make informed calls, divide the labor among your tools clearly first:
- For everyday Q&A, standard models are plenty
- For retrieving new information, use Research tools that can fetch full text
- For high-stakes content, rely on the double safety net of “model summary + manual click-through”
- When dealing with content like X, WeChat Official Accounts, or community forum replies, default to doubting whether the scrape was complete
Here’s another practical rule: if an answer doesn’t cite specific source sentences, or if you can’t click through to verify them, treat that response as a rough draft at best, not a conclusion.
A lot of what gets labeled as “the model getting dumber” is simply starving it of enough information in the first place. Fix retrieval first; then we can talk about boosting intelligence.