AI agents hallucinate prices: LLM grounding with live data

In tests on 21 live product pages from Amazon, Walmart, Best Buy, Target and four Shopify stores, GPT-6 Luna and Claude Opus 5.5 each returned the exact live price for only 2 of 15 products when answering from memory, with median errors of 12% and 17% respectively, according to an Apify analysis. Adding built-in web search raised the exact-match rate but every run still contained wrong prices, each carrying a source link, and a request from outside the US returned an Amazon page in Indian rupees (INR 7,503.02 for an Instant Pot that cost $79 in the US). Apify's proposed fix is a grounded Python agent that pulls live prices from its E-commerce Scraping Tool Actor and lets code, not the model, write each price into the answer, with an explicit failure status when a check fails.

Ask an AI agent with no tools what a Kindle Paperwhite costs, and it can answer $159.99, the old launch price. The live price at the time of testing was $199.99. In tests, GPT-6 Luna and Claude Opus 5.5 each matched the live price exactly for 2 of 15 products from memory. With built-in web search, they gave more exact prices, but every run still had wrong prices, each with a source link. Large language model LLM grounding is the standard way to stop both kinds of wrong price, from memory and from web search. A grounded agent answers only from data that it fetched for the question, and every answer shows where that data came from. For prices, a grounded agent needs live data, a clear status when a check fails, and code, not the model, that writes each price into the answer. The Python agent below has each of these. It gets live prices from E-commerce Scraping Tool https://apify.com/apify/e-commerce-scraping-tool , a ready-to-run tool on the Apify marketplace. Apify calls these tools Actors. First, the tests show where wrong prices come from. Then you build the agent and test it step by step. Why AI agents hallucinate prices If your agent shows prices, a wrong one can reach your user in 3 main ways. The model answers from memory, it reads a bad page, or a tool says "ok" when it failed. The tests below check the first 2, on 21 live product pages from Amazon, Walmart, Best Buy, Target, and 4 Shopify stores. A research study tests the third. The live prices come from E-commerce Scraping Tool, and an independent check with a separate scraping service, from US IP addresses, matched the Actor on all 21 pages. The results include only the 15 products with a single clear price for a new item, as a US shopper sees it without member discounts or coupons. A price is exact only when it matches the live price to the cent. The other 6 pages showed no price, or an unclear seller or size. The 15 prices didn't change while the models answered. From memory An LLM learns prices during training, and many of those prices change before a user asks. With no tools, gpt-6-luna and Claude Opus 5.5 gave these results. The last row adds the prompt "If you cannot know the current price, answer UNKNOWN": | Result | gpt‑6‑luna | Claude Opus 5.5 | |---|---|---| | Answers that included a price | 20 of 21 | 21 of 21 | | Exact prices | 2 of 15 | 2 of 15 | | Median error | 12% | 17% | | Prices given for products whose store page showed no price | 2 of 3 | 3 of 3 | | UNKNOWN answers with that prompt | 21 of 21 | 21 of 21 | Without the UNKNOWN prompt, Claude Opus 5.5 added a clear warning, such as "I can't check live prices," to all 21 answers. Each answer still included a price, and agent code reads the number, not the warning. So a prompt alone gave the user either a guess or nothing. Prices can also change fast. The AirPods Pro 3 at Best Buy went from $199.99 to $249.99 in the 3 days before the models were asked. Even a price from last week can mislead your user. From a bad page When your agent fetches pages itself, the page can be the wrong one, even when it loads normally. - Wrong country. A request from outside the US received an Amazon page in Indian rupees. The model read it correctly and answered INR 7,503.02 for an Instant Pot that cost $79 in the US. - Wrong seller. An Amazon page showed "No featured offers available." The model answered $71.89. That price came from the line "Other sellers on Amazon New & Used 2 from $71.89," not from the Buy Box, the main offer on an Amazon page. - Wrong variant. Target redirected the link for the black Kindle Paperwhite to its parent product page, and Owala redirected its FreeSip link to the 32 oz size. A price from the new page can belong to another variant. Code can detect each of these cases, because the data shows the currency, the seller, and the final page URL. From a tool that says "ok" when it failed The preprint Fabrication after tool failure https://arxiv.org/abs/2609.14758 , a study shared before peer review, tested what models do when a tool returns bad data. When the tool returned "ok" with bad data, such as an empty, outdated, or incomplete result, up to 45.3% of answers were dishonest. A dishonest answer hid the failure. When the tool returned an error status instead, the study reported 0% dishonest answers. So a price tool that reports a clear status or error helps the agent give honest answers. What the 3 causes have in common In all 3 causes, the wrong price looks like a normal answer, and nothing tells the user that something failed. The scraper tests show this. In 3 test runs, a simple do-it-yourself DIY scraper received 26 block pages. A block page is what a store shows when it blocks a request, such as Walmart's "Robot or human?" page or Amazon's "Continue shopping" page. The model answered "no price" for all 26. Those answers were honest, but the user still got no price. Every clearly wrong price in these tests came from a page that looked normal. So live data alone isn't enough. The agent also needs a clear status for every price, so that code can flag a wrong page even when it looks normal. Built-in web search adds live data, but it doesn't give a clear status or code that writes each price into the answer. Way 1: price extraction without Apify Agent builders usually start with one of two options: the model's built-in web search or a scraper they write themselves. With the model's built-in web search The model provider runs the search, and your code only enables 1 setting. In this setup, each model answered 3 times for each of the same 15 products: | Result across 3 runs | gpt‑6‑luna | Claude Opus 5.5 | |---|---|---| | Exact prices | 9 to 11 of 15 | 7 to 9 of 15 | | Wrong prices | 4 to 5 of 15 | 5 to 8 of 15 | | No price | 0 to 1 of 15 | 0 to 1 of 15 | | Biggest price error in a run | 25% to 50% | 25% to 30% | Web search gave many more exact prices than memory, but every wrong price included a source link. Those links went to price tracker sites, store search pages, and pages for other products. This was the answer from Claude Opus 5.5 for the Dyson V15 Detect at Best Buy, where the live price was $899.99: {"price": 629.99, "currency": "USD", "source url": "https://www.bestbuy.com/product/dyson-v15-detect-extra-cordless-vacuum-with-10-accessories-yellow-nickel/J3ZCSY279C"} The link opens a different product, the V15 Detect Extra. If your agent uses web search for prices, a wrong price looks the same as a correct one, because each answer cites a source. With your own scraper The DIY scraper fetched each product page and read the schema.org offers object, which many stores add to the page as JSON-LD structured data for search engines . When a page had no JSON-LD price, gpt-6-luna read the page text instead. The same scraper ran from a US cloud server and from a computer outside the US, with 2 HTTP clients. httpx is a standard Python client, and curl cffi copies the TLS fingerprint of a real Chrome browser. The last row shows the better of the 2 clients for each location: | Store | From a US cloud server | From outside the US | |---|---|---| | Amazon 5 pages | Mostly the "Continue shopping" block page | Pages in rupees, or the block page | | Walmart 3 pages | "Robot or human?" with both clients | Product pages with httpx , "Robot or human?" with curl cffi | | Best Buy 5 pages | Timeouts and dropped connections | Timeouts with httpx , prices with curl cffi | | Target 4 pages | Pages loaded, but the price wasn't in the HTML | Pages loaded, but the price wasn't in the HTML | | Shopify stores 4 pages | Prices for Ridge and Allbirds | Prices for Allbirds, and for Ridge with curl cffi | | Exact prices | 3 of 15 | 6 of 15 | These are 2 of the pages that the DIY scraper received: Patchright is an open-source version of Playwright that hides browser automation. With Patchright and Chrome, the scraper got the exact price for 7 of 15 products from outside the US. Walmart still showed the "Robot or human?" page, and Amazon and most Target pages showed no US price. All these results come from basic scrapers with no proxies. A production scraper needs US residential proxies and code for each store, which means more cost and maintenance for your team. Experienced scraper builders often try other sources before a browser. These include the site's internal API, the /products/<handle .js endpoint on many Shopify stores, and official APIs such as the Best Buy Products API. Each works for only some stores. In the test, the Shopify endpoint worked on 2 of the 4 Shopify stores. For the basics of DIY scraping, see how to scrape prices from websites with Python https://blog.apify.com/price-scraping-python/ . Way 2: price extraction with E-commerce Scraping Tool E-commerce Scraping Tool https://apify.com/apify/e-commerce-scraping-tool apify/e-commerce-scraping-tool gets product data, including prices, from product URLs on Amazon, Walmart, Best Buy, Target, Shopify stores, and many more. It returns the same output format for every store, with fields such as offers.price , offers.priceCurrency , url , and scrapedAt . It manages proxies and browser rendering itself, so your code doesn't need extra logic for block pages or for getting the US version of a page. With scrapeMode set to AUTO , it chooses plain HTTP or a full browser for each page. With additionalProperties set to true, it also returns the seller and the condition of the Buy Box offer when the store shows them. The condition tells you whether the offer is new or used. Install the Apify API client for Python https://docs.apify.com/api/client/python/docs : pip install apify-client==3.3.0 Then set APIFY TOKEN in your environment and call the Actor with your product URLs: python import os from apify client import ApifyClient urls = "https://www.amazon.com/dp/B0F3PT1VBL" client = ApifyClient os.environ "APIFY TOKEN" run = client.actor "apify/e-commerce-scraping-tool" .call run input={ "detailsUrls": {"url": u} for u in urls , "scrapeMode": "AUTO", "additionalProperties": True, } for item in client.dataset run.default dataset id .iterate items : print item "inputUrl" , item.get "offers" , item "scrapedAt" The output: https://www.amazon.com/dp/B0F3PT1VBL {'price': 458, 'priceCurrency': '$'} 2026-10-24T04:24:36.009Z This is the same run in Apify Console: In test runs on 4 different days, the Actor returned a price or a clear "no price" for every one of the 21 pages. This included the Amazon, Walmart, and Best Buy pages where the DIY scraper received block pages or timeouts. With a clear "no price," the agent can say "not available" instead of guessing. Where a store showed no price, such as a sold-out Best Buy product, the independent check also found none. The Apify Free plan includes $5 of usage every month, enough for hundreds of live price checks. The Actor's pricing page https://apify.com/apify/e-commerce-scraping-tool/pricing shows the current prices on each plan. For live price checks, start a separate Actor run for each product, and run them in parallel. Your agent then receives each price as soon as its own run finishes. To fill a cache, 1 run is enough for the whole catalog. Web search, DIY scraper, Shopify, store APIs, and Apify compared The tables below compare 5 price sources for an agent, 1 in each row. The Shopify row covers 2 options. One is the /products/<handle .js endpoint from Way 1. The other is Shopify's endpoint for the Model Context Protocol MCP , called Shop MCP. In the tests, Shop MCP returned products from each Shopify store's catalog, and the agent had to match the exact variant itself. What each source covers and returns: | Price source | Stores | Output | Pages with no price | |---|---|---|---| | Built-in web search | Any page the search finds | A text answer with a source link | Can return a price from another page | | DIY scraper | Pages that accept your requests | Different for every store | Your code must check that it isn't a block page | | Shopify endpoint or Shop MCP | Shopify stores | Shopify format, and the agent must match the variant | Shows availability | | Official store APIs | 1 store per API | Different for every API | Depends on the API | | E-commerce Scraping Tool | Amazon, Walmart, Best Buy, Target, Shopify stores, and many more | Same format for every store | Clear "no price" | What each source takes to run: | Price source | Scraper maintenance | Cost when tested | |---|---|---| | Built-in web search | None | The provider's search fees | | DIY scraper | You update it when a store changes | Your time, plus proxies in production | | Shopify endpoint or Shop MCP | Low | Free | | Official store APIs | Low | Free, with terms of use | | E-commerce Scraping Tool | None Apify maintains the Actor | Free plan with $5 of usage every month, then usage-based | The agent below uses E-commerce Scraping Tool because it returns the same format for every store and a clear "no price" when a store shows none. Build a grounded price agent The agent uses 4 files, catalog.json , prices.py , agent.py , and refresh.py . Step 4 adds a test script, eval short.py . This is how a price moves from the store page to the answer: Before you start You need Python 3.11 to 3.14, an Apify account https://console.apify.com/sign-up the Free plan needs no credit card , and an OpenAI API key. Install the packages: pip install apify-client==3.3.0 openai==3.35.1 python-dotenv==1.1.1 Find your Apify token https://docs.apify.com/integrations/api in Apify Console. Then create a project folder with a .env file: APIFY TOKEN=your apify token OPENAI API KEY=your openai key OPENAI MODEL=gpt-6-luna Add a catalog.json file with the products your agent can check. For a product with colors or sizes, link to the exact variant, because a product URL without one usually shows the default variant. E-commerce Scraping Tool can also help you find product URLs. It searches by keyword on the marketplaces you choose. Search results often list each model and color separately, so you can choose the exact one for your catalog. This catalog has the 7 entries that the examples in Step 3 use. Add your own products the same way: {"store": "Amazon", "product": "Sony WH-1000XM6 headphones", "url": "https://www.amazon.com/dp/B0F3PT1VBL", "currency": "USD"}, {"store": "Walmart", "product": "Sony WH-1000XM6 headphones", "url": "https://www.walmart.com/ip/Sony-WH1000XM6-Noise-Canceling-Wireless-Over-Ear-Headphones/16504921979", "currency": "USD"}, {"store": "Best Buy", "product": "Sony WH-1000XM6 headphones", "url": "https://www.bestbuy.com/product/sony-wh-1000xm6-best-wireless-noise-cancelling-headphones-black/J7XSRH5RCF", "currency": "USD"}, {"store": "Target", "product": "Sony WH-1000XM6 headphones", "url": "https://www.target.com/p/sony-wireless-noise-canceling-headphones-black/-/A-94649372", "currency": "USD"}, {"store": "Walmart", "product": "Instant Pot Duo 7-in-1 6 qt", "url": "https://www.walmart.com/ip/Instant-Pot-DUO60-V4-6-Quart-Duo-Electric-Pressure-Cooker-Slow-Cooker/45918917", "currency": "USD"}, {"store": "Best Buy", "product": "Instant Pot Duo 7-in-1 6 qt", "url": "https://www.bestbuy.com/product/instant-pot-6-quart-duo-7-in-1-electric-pressure-cooker-silver-brushed-stainless-steel/J3RK3ZJ866", "currency": "USD"}, {"store": "Amazon", "product": "Dyson V15 Detect cordless vacuum", "url": "https://www.amazon.com/dp/B0C2J8KJH9", "currency": "USD"} By the end of Step 4, the project folder looks like this: price-agent/ ├── .env ├── catalog.json ├── prices.py ├── agent.py ├── refresh.py ├── eval short.py └── price cache.json Step 1. Build the price tool prices.py fetches prices from E-commerce Scraping Tool, gives every product a status, and names the seller when the Buy Box offer comes from a seller other than the store. The status flags cases such as a sold-out item, another currency, or a redirect, even when the page looks normal. Each status comes from fields that the Actor returns, such as the price, the final page URL, the seller, and the condition. - ok means the price came from the page at the catalog URL. - no price means the page showed no price. - other currency means the price isn't in the currency that the catalog lists for that store. - used means the Buy Box offer is for a used item, such as "Used - Like New." - moved page means the store redirected the request to a different page, such as a parent product page. - big change means the price moved more than 30% from its reference price, the last ok price in the cache. This can be a normal price change, or a different seller, size, or condition, so the agent asks before it acts. The flag stays for 1 day, so every check during that day reports it. After that, the new price becomes the reference. - not live means the answer uses the last known price and its time instead of a live result. Set the change threshold for your products. In the tests, a 30% threshold flagged the Amazon Dyson listing when its price fell from $789.49 to $377.99 within 3 days. The same threshold didn't flag a normal 25% price change on AirPods at Best Buy, a few days before the models answered. Browsing questions use a cached price up to 1 hour old. Decisions, such as a purchase or a price alert, get a live price check, because the prompt tells the model to set need live to true for them. In your own checkout code, call asyncio.run get prices products, need live=True directly, so the live price check doesn't depend on the model. Set TIME LIMIT to the longest wait your users accept. A check that takes longer returns the last known price, if there is one, with the not live status. A failed run, or a run with no result for the page, works the same way, so a failed check never becomes a new no price row. The file has 5 parts: - make row converts 1 Actor result into a row with a status. - check live sends 1 product to the Actor and waits up to TIME LIMIT for the result. - compare flags a big price change and keeps the flag for 1 day. - get prices uses the cache for browsing and a live price check for decisions. - refresh all and load from task fill the cache from 1 run or from a scheduled task. This is prices.py : python import asyncio import json import os import re from datetime import datetime, timedelta, timezone from pathlib import Path from urllib.parse import urlparse from apify client import ApifyClientAsync ACTOR = "apify/e-commerce-scraping-tool" CACHE FILE = Path "price cache.json" TIME LIMIT = timedelta seconds=60 longest wait for a live price check before an action MAX PARALLEL = 4 live price checks that run at the same time MAX AGE = timedelta hours=1 browsing answers may use prices up to this old BIG CHANGE = 0.3 a price that moved more than 30% from the reference price HOLD = timedelta days=1 a big change stays flagged this long, then becomes the new reference DOLLAR = { ".ca": "CAD", ".com.au": "AUD", ".com.mx": "MXN", } "$" on these sites isn't USD; add the domains in your catalog def to number value : """Return the price as a number.""" if value in None, "" : return None text = str value if re.search r",\d{2}$", text : "1.299,00" style text = text.replace ".", "" .replace ",", "." match = re.search r"\d \d, \.?\d ", text return float match.group .replace ",", "" if match else None def to currency value, expected, url : """Return the currency code, with the catalog currency as the default.""" if value in None, "" : return expected if value == "$": host = urlparse url .netloc return next code for end, code in DOLLAR.items if host.endswith end , "USD" return "USD" if value in "US$", "USD" else value def seller and condition item : """Return the seller and condition of the Buy Box offer.""" extra = item.get "additionalProperties" or {} main = next s for s in extra.get "sellerOffers" or if s.get "isBuyBoxWinner" , {} seller = main.get "sellerName" or extra.get "seller" if isinstance seller, dict : seller = seller.get "name" return seller, main.get "condition" def make row item, product : offer = item.get "offers" or {} price = to number offer.get "price" currency = to currency offer.get "priceCurrency" , product "currency" , item.get "url" or product "url" , same page = item.get "url", "" .split "?" 0 .rstrip "/" == product "url" .split "?" 0 .rstrip "/" seller, condition = seller and condition item if price is None: status = "no price" elif currency = product "currency" : status = "other currency" elif condition and not condition.lower .startswith "new" : status = "used" elif not same page: status = "moved page" else: status = "ok" return { "store": product "store" , "product": product "product" , "url": product "url" , "price": price, "currency": currency, "seller": seller, "condition": condition, "checked at": item.get "scrapedAt" , "status": status, } async def check live client, limit, product : """Send one product to the Actor and wait up to TIME LIMIT for its result.""" async with limit: run = await client.actor ACTOR .start run input={ "detailsUrls": {"url": product "url" } , "scrapeMode": "AUTO", "additionalProperties": True, } adds the seller and condition finished = await client.run run.id .wait for finish wait duration=TIME LIMIT if finished is None or finished.status in "READY", "RUNNING" : await client.run run.id .abort end the run when the agent stops waiting return None if finished.status = "SUCCEEDED": return None run = finished items = i async for i in client.dataset run.default dataset id .iterate items return make row items 0 , product if items else None def compare row, old : """Mark a price that moved a lot from the reference price. It can be a real price drop, or a different seller, size, or condition, so the agent asks before acting.""" old = old or {} ref = old.get "reference price" or old.get "price" if old.get "status" == "ok" else None if ref and row "price" and row "status" == "ok" and abs row "price" - ref / ref BIG CHANGE : since = old.get "changed since" or row "checked at" if age {"checked at": since} < HOLD: refreshes keep the flag for a day return { row, "status": "big change", "previous price": ref, "reference price": ref, "changed since": since, } return { row, "reference price": row "price" if row "status" == "ok" else ref} def load cache : return json.loads CACHE FILE.read text if CACHE FILE.exists else {} def age row : if not row.get "checked at" : return timedelta.max checked = datetime.fromisoformat row "checked at" .replace "Z", "+00:00" return datetime.now timezone.utc - checked async def get prices products, need live : """Browsing questions use a cached price up to MAX AGE old. Calls with need live=True an alert, a checkout, an approval always get a live price check. """ client = ApifyClientAsync os.environ "APIFY TOKEN" limit = asyncio.Semaphore MAX PARALLEL cache = load cache rows, to check = {}, for product in products: cached = cache.get product "url" if not need live and cached and age cached <= MAX AGE: rows product "url" = cached else: to check.append product results = await asyncio.gather check live client, limit, p for p in to check for product, row in zip to check, results, strict=True : if row: row = compare row, cache.get product "url" cache product "url" = row rows product "url" = row elif product "url" in cache: Use the last known price, with its time, and mark it so the agent says it's not live. cached = cache product "url" status = "not live" if cached "status" in "ok", "moved page", "big change", "used" else cached "status" rows product "url" = { cached, "status": status} else: rows product "url" = { product, "price": None, "checked at": None, "status": "not live", } CACHE FILE.write text json.dumps cache, indent=1 return rows p "url" for p in products async def save items dataset, products : """Convert a dataset of Actor results into cache rows.""" by url = {p "url" : p for p in products} cache = load cache async for item in dataset.iterate items : product = by url.get item.get "inputUrl" if product: cache product "url" = compare make row item, product , cache.get product "url" CACHE FILE.write text json.dumps cache, indent=1 return cache async def refresh all products : """Fill the cache for every product in one run. Run it every 30 minutes.""" client = ApifyClientAsync os.environ "APIFY TOKEN" run = await client.actor ACTOR .call run input={ "detailsUrls": {"url": p "url" } for p in products , "scrapeMode": "AUTO", "additionalProperties": True, } return await save items client.dataset run.default dataset id , products async def load from task task id, products : """Fill the cache from the last successful run of a scheduled Apify task.""" client = ApifyClientAsync os.environ "APIFY TOKEN" last run = client.task task id .last run status="SUCCEEDED" return await save items last run.dataset , products Step 2. Build an agent where code inserts each price A common grounding check compares the numbers in the answer with the numbers in the tool output. The test catalog was larger than the 7 entries above and included the Kindle Paperwhite. These 5 wrong answers use only prices from that tool output, such as $199.99, the Kindle's price. This grounding check passes all of them, and the validator in the agent below rejects all of them: | Wrong answer | Example | Grounding check | This agent's validator | |---|---|---|---| | Swapped stores | "The cheapest is Target at $458.00. Amazon charges $458.99." | Passes | Rejects | | Price on the wrong product | "The Kindle Paperwhite costs $458.00 at Amazon." | Passes | Rejects | | Price for a sold-out item | "The Instant Pot is $199.99 at Best Buy." | Passes | Rejects | | Price in words | "The Sony headphones cost four hundred fifty-eight dollars at Walmart." | Passes | Rejects | | Price in euros | "At Amazon, the Sony WH-1000XM6 is EUR 458." | Passes | Rejects | The check proves that a number exists in the data. It doesn't prove that the number belongs to that product and that store. A source link has the same weakness, because a user can only check it after reading the answer. So in the agent below, the model writes placeholders, code inserts each price from the data, and the validator checks the answer before the user sees it. Apify CEO Jan Čurn describes the same approach in his talk at the Prague Crawl conference, MCP doesn't suck, your agent does https://www.youtube.com/watch?v=MnbXQr1UQoM . agent.py uses the OpenAI Responses API with 1 tool, get prices . The placeholders look like {{offer:r1}} or {{cheapest}} . For {{cheapest}} , code chooses the cheapest offer among rows with the status ok . The cheapest offer is based on listed prices. Before a purchase, take the final total, with shipping and tax, from the store's checkout. Rows with no price, or with a price in another currency, get no placeholder. Code adds them to the end of the answer, together with any checked offer that the model skipped. The tool schema lists the catalog URLs as an enum , so the model can choose only those URLs. The validator rejects an answer when any of these is true: - The model typed a currency sign or code. - The model typed a price with cents. - The model typed a capitalized store name, such as "Target." - The answer names a product with no offer in it. - The answer calls an offer the lowest without {{cheapest}} . Each inserted offer names its product, so the reader sees which product each price belongs to. The validator returns the error to the model up to 2 times. After that, the agent returns a short fallback message. The validator doesn't check plain numbers with no currency sign, such as "458." For those, the prompt adds a second check and tells the model not to write prices. The file has 4 parts: - SYSTEM tells the model to write placeholders and never a price or a store name. - run tool calls get prices and gives each row a row id . - render runs the validator checks, then replaces each placeholder with data from the rows. - ask runs the loop and gives the model 2 retries after a rejected draft. This is agent.py : """A price agent in which code inserts each price. The model writes {{offer:r1}}, and code inserts the price, store, time, and status from live data.""" import asyncio import json import os import re import sys from datetime import datetime from dotenv import load dotenv load dotenv from openai import OpenAI noqa: E402 from prices import get prices noqa: E402 MODEL = os.environ.get "OPENAI MODEL", "gpt-6-luna" set OPENAI MODEL in .env to use another model MAX TURNS = 8 stop a model that keeps calling tools FALLBACK = "Sorry, I could not answer that. Please ask again." CATALOG = json.load open "catalog.json" A brand store, such as Ridge for a Ridge wallet, shares its name with the product, so the model may write it. Every other store name must come from code. STORES = sorted { p "store" for p in CATALOG if not any p "store" .lower in q "product" .lower for q in CATALOG } SYMBOL = {"USD": "$", "EUR": "€", "GBP": "£"} oa = OpenAI TOOLS = { "type": "function", "name": "get prices", "description": "Get prices for products from the catalog. Each row has a row id and a status.", "parameters": { "type": "object", "properties": { The enum lets the model choose only catalog URLs, so a typo can't create a new one. "urls": { "type": "array", "items": {"type": "string", "enum": p "url" for p in CATALOG }, }, "need live": { "type": "boolean", "description": "true before a purchase, a price alert, or an approval. false for browsing questions.", }, }, "required": "urls", "need live" , "additionalProperties": False, JSON Schema rule: no other arguments }, "strict": True, } SYSTEM = "You answer questions about product prices in one or two full sentences, starting " "with the answer. Never write a price, a currency, " "or a store name yourself. Refer to each offer only as {{offer:ROW ID}}, using " "row id values from get prices, and use a placeholder for every row the user asked " "about. Code fills in the price, product, store, and time. Each placeholder becomes a full phrase, " "such as '$458.00 for the Sony WH-1000XM6 headphones at Amazon checked ... ', so do not " "repeat the price, product, or store it names, and use each " "placeholder only once. Rows with no usable price have no row id. Code adds them to the " "answer, so do not mention them. If no row has a usable price, write one short sentence " "saying that it has no listed price right now. " "To name the cheapest offer for one product, write {{cheapest}} and code fills in every offer at " "the lowest price. " "Set need live to true only before a purchase, a price alert, or an approval. Words like " "'today' or 'now' in a question do not make it a purchase. If a status is not live and the " "user wants to act, say you could not confirm the live price. If a status is big change, " "give the offer first, then ask the user to confirm the new price before acting. " "If a status is used, say that the offer is for a used item. " "If a product is not in the catalog, say you cannot check it.\nCatalog:\n" + json.dumps {k: p k for k in "store", "product", "url" } for p in CATALOG ANSWER FORMAT = { "type": "json schema", "name": "answer", "strict": True, "schema": { "type": "object", "properties": {"text": {"type": "string"}}, "required": "text" , "additionalProperties": False, }, } PLACEHOLDER = re.compile r"\{\{offer: r\d+ \}\}|\{\{cheapest\}\}" MONEY = re.compile r" $€£₹ |\b USD|EUR|GBP|INR|dollars?|euros? \b|\b\d+\.\d{2}\b", re.IGNORECASE Words that name each catalog product, such as "sony" or "kindle". PRODUCT WORDS = { p "product" : { w for w in re.findall r" \w- +", p "product" .lower :2 if len w 3 } for p in CATALOG } SUPERLATIVE = re.compile r"\b cheapest|lowest|least expensive|best price \b", re.IGNORECASE STORE NAME = re.compile r"\b " + "|".join map re.escape, STORES + r" \b" capitalized store names only STATUS NOTE = { "ok": "", "moved page": ", the store redirected this link", "not live": ", last known price, not live", } def typed fact text : """Find a currency sign or code, a price with cents, or a capitalized store name the model typed.""" text = PLACEHOLDER.sub "", text return MONEY.search text or STORE NAME.search text def join parts : return parts 0 if len parts == 1 else ", ".join parts :-1 + ", and " + parts -1 rows = {} cheapest = unpriced = rows with no usable price, which code adds to the answer def run tool urls, need live : products = p for p in CATALOG if p "url" in urls found = asyncio.run get prices products, need live out = for row in found: if row "price" is None or row "status" == "other currency": if row not in unpriced: unpriced.append row out.append { "store": row "store" , "product": row "product" , "status": row "status" , "price": None, "note": "code adds this row to the answer", } continue row id = f"r{len rows + 1}" rows row id = row out.append { "row id": row id, "store": row "store" , "product": row "product" , "status": row "status" , "price": row "price" , } The cheapest offer covers every row in this answer, not only this tool call. priced = { rid: r for rid, r in rows.items if r "price" is not None and r "status" == "ok" } low = min priced.values , key=lambda r: r "price" "price" if priced else None cheapest : = rid for rid, r in priced.items if r "price" == low return {"rows": out} def describe row : """Convert a row into the phrase that replaces its placeholder.""" if row "price" is None and row "status" == "not live": return f"Price not confirmed for the {row 'product' } at {row 'store' }" if row "price" is None or row "status" == "no price": return f"{row 'store' } shows no price for the {row 'product' }" if row "status" == "other currency": return f"No price in the expected currency for the {row 'product' } at {row 'store' }" checked = datetime.fromisoformat row "checked at" .replace "Z", "+00:00" sign = SYMBOL.get row "currency" , row "currency" + " " price = f"{sign}{row 'price' :,.2f}" note = STATUS NOTE.get row "status" , "" if row "status" == "big change": note = f", was {sign}{row 'previous price' :,.2f} before the change" if row "status" == "used" or row "status" == "not live" and row.get "condition" and not row "condition" .lower .startswith "new" : note += f", {row 'condition' }" seller = row.get "seller" if seller and row "store" .lower not in seller.lower : note = f", sold by {seller}" + note another seller on the store's site return f"{price} for the {row 'product' } at {row 'store' } checked {checked:%H:%M} UTC on {checked:%B} {checked.day}{note} " def render text : """Reject any price or store name the model typed. Fill every placeholder from data.""" if typed fact text : return None, "You typed a price, currency, or store name. Use only {{offer:ROW ID}} placeholders.", if SUPERLATIVE.search PLACEHOLDER.sub "", text and "{{cheapest}}" not in text: return None, "To name the lowest price, use {{cheapest}}, which code fills in." if "{{cheapest}}" in text and len {r "product" for r in rows.values } 1: return None, "Use {{cheapest}} only when the answer covers one product. Use {{offer:ROW ID}} instead.", ids = i for i in PLACEHOLDER.findall text if i + cheapest if "{{cheapest}}" in text else offered = {rows i "product" for i in ids if i in rows} | { r "product" for r in unpriced } words = set re.findall r" \w- +", PLACEHOLDER.sub "", text .lower if any w & words for p, w in PRODUCT WORDS.items if p not in offered : return None, "You named a product that has no offer in your answer. Name only the products of the offers you use.", used = set def fill match : if match.group 1 is None: {{cheapest}} if not cheapest: return "no confirmed price" used.update cheapest return join describe rows r for r in cheapest row = rows.get match.group 1 if row is None: raise KeyError match.group 1 used.add match.group 1 return describe row try: answer = PLACEHOLDER.sub fill, text .strip others = describe r for rid, r in rows.items if rid not in used offers the model skipped if unpriced or others and answer and answer -1 not in ". ?": answer += "." lead = "Another offer is" if len others == 1 else "Other offers are" extra = f"{lead} {join others }." if others else answer = " ".join answer :1 .upper + answer 1: + extra + describe r + "." for r in unpriced return answer.strip , None except KeyError as missing: return None, f"Row {missing} does not exist. Use only row id values from get prices.", def ask question, max retries=2 : rows.clear cheapest.clear unpriced.clear items = {"role": "user", "content": question} response = oa.responses.create model=MODEL, instructions=SYSTEM, input=items, tools=TOOLS, text={"format": ANSWER FORMAT}, retries = 0 for in range MAX TURNS : calls = o for o in response.output if o.type == "function call" if calls: items += response.output for call in calls: args = json.loads call.arguments result = run tool args "urls" , args "need live" items.append { "type": "function call output", "call id": call.call id, "output": json.dumps result , } else: answer, error = render json.loads response.output text "text" if answer: return answer if retries == max retries: return FALLBACK retries += 1 items += response.output + {"role": "user", "content": error} response = oa.responses.create model=MODEL, instructions=SYSTEM, input=items, tools=TOOLS, text={"format": ANSWER FORMAT}, return FALLBACK if name == " main ": print ask " ".join sys.argv 1: or "Where is the Sony WH-1000XM6 cheapest right now?" Step 3. Fill the cache and ask questions refresh.py fills the cache for the whole catalog in 1 run: python """Fill the price cache for the whole catalog. Run it on a schedule.""" import asyncio import json from dotenv import load dotenv load dotenv from prices import refresh all noqa: E402 catalog = json.load open "catalog.json" cache = asyncio.run refresh all catalog print f"Cached {len cache } products" Run it, then ask the agent a browsing question: python refresh.py python agent.py "Where is the Sony WH-1000XM6 cheapest right now: Amazon, Walmart, Best Buy, or Target?" The output: Cached 7 products The lowest-priced offers are tied: $458.00 for the Sony WH-1000XM6 headphones at Amazon checked 05:25 UTC on October 24 , $458.00 for the Sony WH-1000XM6 headphones at Walmart checked 05:26 UTC on October 24 , and $458.00 for the Sony WH-1000XM6 headphones at Best Buy checked 05:26 UTC on October 24 . Another offer is $458.99 for the Sony WH-1000XM6 headphones at Target checked 05:25 UTC on October 24, the store redirected this link . The answer used cached prices, so it needed no live price check. Target's offer has a note because Target redirected the catalog link to its parent product page. For a product with no price on the store's page, the agent says so. Best Buy showed this Instant Pot as sold out: python agent.py "How much is the Instant Pot Duo 6 qt at Best Buy?" The output: It has no listed price right now. Best Buy shows no price for the Instant Pot Duo 7-in-1 6 qt. A question before a purchase gets a live price check. A sudden low price needs extra care. During the tests, the Amazon listing for the Dyson V15 Detect went from $789.49 to $377.99, then to $422.44: python agent.py "I want to buy the Dyson V15 Detect at Amazon now. What does it cost?" The output: It's $422.44 for the Dyson V15 Detect cordless vacuum at Amazon checked 05:26 UTC on October 24, sold by Evergreen Renew Tech LLC, Used - Like New ; note that this offer is for a used item. This is the Amazon page a US shopper saw at the time of that check: E-commerce Scraping Tool reported the seller and the condition of the Buy Box offer, and code used them to set the used status. The status is used , not big change , because the code checks the condition before it compares prices. A grounding check would pass "$422.44 at Amazon" with no used note, because that number is in the data. If a setup step is missing, the error message names it: | Error | Fix | |---|---| | KeyError: 'APIFY TOKEN' | Create .env in the project folder, next to the scripts. | | openai.OpenAIError: Missing credentials | Add OPENAI API KEY to .env . | | AuthenticationError or UnauthorizedError | Copy the key or token into .env again, without spaces or quotes. | | The answer says a product isn't in the catalog | Add the product to catalog.json with its exact product URL. | Step 4. Test the agent with repeated questions A single good answer isn't enough proof, because a model can answer differently each time. Ask each test question several times, and check what the agent did, not only what it said. This script, eval short.py , checks which products each answer used: """Ask each question 5 times and check which products the agent used.""" import agent TESTS = { "Where is the Sony WH-1000XM6 cheapest right now: Amazon, Walmart, Best Buy, or Target?": "Sony", "How much is the Instant Pot Duo 6 qt at Best Buy?": "Instant Pot", "How much is the Nintendo Switch 2 at Walmart?": None, not in the catalog } for question, product in TESTS.items : for in range 5 : answer = agent.ask question used = {row "product" for row in list agent.rows.values + agent.unpriced} right = all product in p for p in used and bool used if product else not used print "PASS" if right and answer = agent.FALLBACK else "FAIL", "|", answer :90 Run it with python eval short.py . In the tests, 6 runs of this script passed 89 of 90 checks. A longer run asked 8 questions, 5 times each. The questions included a purchase, a price alert, and a used offer, and the run checked whether the agent asked for a live price check. It gave these results: - All 40 answers used the correct rows and the correct kind of price check, cached or live. - The validator worked as designed. It found a store name that the model typed in 12 first drafts. Each rejected draft passed on the second try. - No answer needed the fallback message. Run the script again after every change to the prompt or the validator, because repeated runs find bugs that a single run misses. Run the price agent in production ChatGPT and the Universal Commerce Protocol UCP https://ucp.dev/specification/overview/ specification from Google use 2 kinds of price. The feed price is for browsing, and the checkout price is the final one. Your agent can work the same way. Each kind of agent needs a different price: | Agent | Typical question | Price it uses | |---|---|---| | Shopping assistant | "Where is it cheapest?" | Cached price, up to 1 hour old | | Price alert | "Tell me when it drops below my target price." | Live price check before the alert is sent | | Purchase or approval agent | "Buy it" or "Approve this order." | Live price check, and a person approves any big price change | - Keep the cache up to date. Save the Actor input from refresh.py as a task in Apify Console https://console.apify.com/ . Run the task every 30 minutes with an Apify schedule https://docs.apify.com/actors/running/schedules , so the refresh runs on Apify, not on your server. Then read its results with load from task , or set an Apify webhook https://docs.apify.com/integrations/webhooks to call your code when the run succeeds. If you already use cron, run refresh.py every 30 minutes instead. - Monitor every status. Log each status separately, such as no price and moved page . A change at one store is then visible, instead of hidden in an overall success rate. Also log the seller and condition for each store, because the used status depends on those fields. To read the results of the task's last successful run, replace the import and the refresh all call in refresh.py with these lines: python from prices import load from task noqa: E402 cache = asyncio.run load from task "YOUR TASK ID", catalog The tutorial code keeps state in 1 process and 1 JSON file. For a web service, keep the rows for each request separate, call get prices with await , and store the cache in a database. For a catalog with thousands of products, don't list every URL in the prompt and the tool schema. Instead, give the model a search tool that returns matching catalog entries. In MCP clients such as Claude or Cursor, the Apify MCP server https://docs.apify.com/mcp makes E-commerce Scraping Tool available to the agent as a tool. Real-time product data for AI agents https://blog.apify.com/real-time-product-data-for-ai-agents/ shows that setup. The case study Maven AGI x Apify: Customer support with 100% company-approved data https://blog.apify.com/maven-agi-customer-support-case-study/ shows grounded agents in production. Next steps Models, prices, and store pages will keep changing, but one rule stays the same. When the model types a price, the user can't know whether it's a guess or a fact. When code inserts it from live data, with a source and a time, the user receives an answer they can check. For more on agents and live data, see AI agent tools that keep agents grounded in current data https://blog.apify.com/ai-agent-tools/ and Agentic commerce and the AI economy stack https://blog.apify.com/agentic-commerce/ . The guide on whether web scraping is legal https://blog.apify.com/is-web-scraping-legal/ answers common legal questions about collecting prices. Every agent needs proper tools. Equip your AI agents with thousands of tools on Apify, starting with E-commerce Scraping Tool https://apify.com/apify/e-commerce-scraping-tool for live prices with a clear status. It runs on the Apify Free plan. Add your own products to catalog.json , and ask the agent where they're cheapest. Try E-commerce Scraping Tool https://console.apify.com/actors/2APbAvDfNDOWXbkWf?addFromActorId=2APbAvDfNDOWXbkWf FAQ What is LLM grounding? LLM grounding means that a model answers from data that your code fetched for the question, not from what it learned in training. A grounded answer shows its source and its time, so a user or a program can check it. Is LLM grounding the same as RAG? RAG retrieval-augmented generation is one way to ground a model. It finds documents and adds them to the prompt. For prices, a retrieved page can be out of date or describe another product, so a price agent should fetch the live product page and let code insert the price. Can a prompt stop price hallucinations? A prompt can stop the guessing. With no tools and the prompt "If you cannot know the current price, answer UNKNOWN," gpt-6-luna and Claude Opus 5.5 both answered UNKNOWN for all 21 products. But then the user got no price. Live data gives an answer that's both honest and useful. Do agentic commerce protocols fix wrong prices? They help at checkout. Protocols such as Google's Universal Commerce Protocol UCP separate a feed price for browsing from a checkout price, which is final. Before checkout, the agent still needs live data with a clear status for every price it shows. How do I get live prices for an AI agent? Give the agent a tool that fetches the product page when the user asks. E-commerce Scraping Tool, a ready-to-run tool on Apify, returns the price, currency, and timestamp for Amazon, Walmart, Best Buy, Target, Shopify stores, and many more. With additionalProperties set to true, it also returns the seller and condition when the store shows them.