Grok step 2, search plus a take

the part I skipped the first week was checking whether the take was even from this year. Grok sounded current. The citation was a recap of a recap. I almost sent a slack about a fee that had already been dropped.

If you just opened the site, back up to getting started with Grok. This step assumes you can get a chat going on grok.com and you have one real question, not a vibe.

What this tool is actually for

xAI's product page says the quiet part. Grok searches the web and X live so the answer is supposed to reflect what is happening now. That is the difference people mean when they say Grok is "search-y." Claude will sit with a long messy draft. Cursor will sit in your files. Grok will go look, then give you a take with some personality on top.

The personality is official too. They want a little wit. Fine. On a news-ish question I care more about the date on the source than the joke, even when its a good joke.

A current take is not the same as a rumor. "What are people on X saying about the outage" is a real ask if you treat it as weather, not as truth. "Did the company change the warranty in 2026" is a better one because there should be a page.

I use Grok for three kinds of questions that keep showing up in regular weeks. What changed since I last looked. Which of two public options is less dumb right now. And the change-fee class of problem, a policy that moves and the company's own site disagrees with itself.

I do not use it as my newspaper. I use it as a first pass when I would have opened eight tabs and then forgotten what I decided.

Ask what changed, with a year on it

"Is this still true" is a lazy prompt and Grok will try anyway. Put the year in. Put the last thing you believed in.

Example I actually typed, more or less:

I last looked at the Remarkable Paper Pro in late 2025 and I thought it was the one to buy. In August 2026 is that still the tablet people mean, or did they ship a replacement. Compare the current Paper Pro to a Kindle Scribe using the official product pages and say what changed in the last year. If a review is from 2025 flag it.

Notice the job. Not "which tablet is best." Best is how you get a shopping essay. You want a delta. What moved.

On a news-ish question I get more specific than "whats going on with X." "What did the company itself post today about the outage, and what are the two most repeated complaints on X that are not from the same three accounts." You will still have to judge the accounts. Grok can search public posts. It cannot tell you who is a bot.

If it answers from memory and does not search, poke it. "Search. I want pages from this month." On the consumer app they tell you you can switch modes for search and reasoning. I do that when a normal chat feels like it is reciting. I do not treat the mode names as magic. The test is whether new links show up.

Two public options, not a secret third thing

Comparing two things Grok can see is a good use. Comparing your private vendor contract to a blog post is how you get a fluent maybe.

"Southwest vs United on same-day change for a domestic coach ticket I bought on the website, August 2026. Use each airline's own policy page. Table the fee, the window, and the one exception each site mentions. No third airline."

I like asking for a table and then refusing to trust the table. I open both URLs. I have caught a swapped column. Grok will apologize and fix it, which is nice, and also proof you needed the click.

When the two options are products, name the SKU if you can. "the 2024 model" vs "the one on the homepage today" matters. If you cannot name it, ask Grok to name what is currently for sale and what is leftover retail, then you go look at the buy page yourself.

Do not ask it to pick for your personality. "Which one should I buy, I like nice things" is how you get a vibe. "I write on the couch for two hours and I already have a Kindle, do I gain anything in 2026" is a question with a spine.

The links are the job

xAI talks about live citations. Good. Use them like a skeptic not a tourist.

I click the ones that claim a number. Fee, date, "as of", anything that would embarrass me in an email. If the citation is a listicle, I say so. "That source is a roundup. Find the primary page or say you cannot."

Sometimes Grok will cite X posts for a policy that should live on a .com. Fun, and wrong kind of evidence. "Posts are fine for what people are complaining about. For the rule I need the company page."

I also ask "what would make this answer stale by next week." That question makes it admit a bill is in committee or a sale ends Friday. Then I put that sentence in my notes, not the whole chat.

If you are the kind of person who forwards answers, forward the URL too. A take without a link is just a confident paragraph. I have been that paragraph. It feels efficient until someone replies "this changed in March."

When the take gets ahead of the search

Grok is faster to have an opinion than some of the stiffer tools. That is the feature and the leak. You will get a sentence that sounds like a conclusion before you have seen the quote.

My fix is rude and short. "Separate facts you found from your take. Facts first, each with a link. Then two sentences of opinion labeled as opinion." It does it. The second half is usually where the joke lives, which is easier to ignore when it is labeled.

If the joke is in the facts, start a new chat. I have had a thread get punchy and then every follow up came with a bit. Funny. Also I was trying to decide whether to keep a subscription.

This is also where people wander into work Grok should not do. A deep private codebase, a careful legal rewrite, a repo it cannot see. That is step 4, and it is a different failure than a bad citation. A bad citation you can catch. A made-up clause in a contract voice you might send.

For now stay public. Two options. A date. Links you open.

Then do not start a fresh chat to write the email. The research is already here. Step 3 is turning this thread into the thing you actually send, in your voice, not Grok's uncle-at-dinner voice.

If you are still deciding whether this should have been Claude or Cursor, which tool when is the honest overlap. I use Grok when the clock on the information matters. I switch when the clock on the sentence matters.